Spaces:
Running
Running
piotr-szleg-bards-ai
commited on
Commit
•
a91a9a0
1
Parent(s):
c952555
2024-03-13 08:24:22 Publish script update
Browse files- data/time_of_day_comparison.csv +177 -177
- data/time_of_day_plots.csv +0 -0
- pipeline/config.py +2 -1
- pipeline/models.py +5 -11
data/time_of_day_comparison.csv
CHANGED
@@ -1,177 +1,177 @@
|
|
1 |
-
model,time_of_day,
|
2 |
-
01-ai Yi Chat (34B),early morning,9.13007491738049,
|
3 |
-
01-ai Yi Chat (34B),morning,
|
4 |
-
01-ai Yi Chat (34B),afternoon,12.255323665086614,
|
5 |
-
01-ai Yi Chat (34B),late afternoon,9.417656523840767,3.
|
6 |
-
01-ai Yi Chat (34B),evening,8.16779062861488,
|
7 |
-
01-ai Yi Chat (34B),late evening,9.680033167203268,
|
8 |
-
01-ai Yi Chat (34B),midnight,9.311747227774726,
|
9 |
-
01-ai Yi Chat (34B),night,9.625520552260966,
|
10 |
-
Chronos Hermes (13B),early morning,8.063542557544396,
|
11 |
-
Chronos Hermes (13B),morning,8.
|
12 |
-
Chronos Hermes (13B),afternoon,6.651120054020601,
|
13 |
-
Chronos Hermes (13B),late afternoon,7.630303209478205,
|
14 |
-
Chronos Hermes (13B),evening,6.468730966250102,2.
|
15 |
-
Chronos Hermes (13B),late evening,7.710561646355523,2.
|
16 |
-
Chronos Hermes (13B),midnight,7.710010987520218,
|
17 |
-
Chronos Hermes (13B),night,
|
18 |
-
Falcon Instruct (7B),early morning,1.2471453740046574,0.
|
19 |
-
Falcon Instruct (7B),morning,1.4012710821060907,0.
|
20 |
-
Falcon Instruct (7B),afternoon,1.5774729509611387,0.
|
21 |
-
Falcon Instruct (7B),late afternoon,1.0884015900748116,0.
|
22 |
-
Falcon Instruct (7B),evening,0.7874892950057983,0.
|
23 |
-
Falcon Instruct (7B),late evening,1.2339241071180864,0.
|
24 |
-
Falcon Instruct (7B),midnight,0.9382020387894068,0.
|
25 |
-
Falcon Instruct (7B),night,1.2997726936862892,0.
|
26 |
-
LLaMA-2 Chat (7B),early morning,2.8209988124230327,
|
27 |
-
LLaMA-2 Chat (7B),morning,2.
|
28 |
-
LLaMA-2 Chat (7B),afternoon,3.5475496424569024,
|
29 |
-
LLaMA-2 Chat (7B),late afternoon,2.7755608558654785,
|
30 |
-
LLaMA-2 Chat (7B),evening,2.9608939344232734,2.
|
31 |
-
LLaMA-2 Chat (7B),late evening,2.869330001913983,
|
32 |
-
LLaMA-2 Chat (7B),midnight,2.8870244533457656,
|
33 |
-
LLaMA-2 Chat (7B),night,4.6455250724417265,
|
34 |
-
Mistral (7B) Instruct v0.2 (Together AI),early morning,3.627201789248306,
|
35 |
-
Mistral (7B) Instruct v0.2 (Together AI),morning,3.
|
36 |
-
Mistral (7B) Instruct v0.2 (Together AI),afternoon,3.255509059895914,
|
37 |
-
Mistral (7B) Instruct v0.2 (Together AI),late afternoon,3.542298033617545,
|
38 |
-
Mistral (7B) Instruct v0.2 (Together AI),evening,5.496347131400273,
|
39 |
-
Mistral (7B) Instruct v0.2 (Together AI),late evening,3.3933188574654713,
|
40 |
-
Mistral (7B) Instruct v0.2 (Together AI),midnight,4.510189438914205,
|
41 |
-
Mistral (7B) Instruct v0.2 (Together AI),night,3.5451678467578573,
|
42 |
-
Mistral-7B-Instruct-v0.2,early morning,4.625752210617065,
|
43 |
-
Mistral-7B-Instruct-v0.2,morning,3.734026002883911,
|
44 |
-
Mistral-7B-Instruct-v0.2,afternoon,3.228973722457886,0.
|
45 |
-
Mistral-7B-Instruct-v0.2,late afternoon,3.2048643112182615,0.
|
46 |
-
Mistral-7B-Instruct-v0.2,evening,3.397640073299408,
|
47 |
-
Mistral-7B-Instruct-v0.2,late evening,
|
48 |
-
Mistral-7B-Instruct-v0.2,midnight,2.6972494465964183,0.
|
49 |
-
Mistral-7B-Instruct-v0.2,night,4.609350740909576,
|
50 |
-
Mixtral-8x7B-Instruct-v0.1,early morning,
|
51 |
-
Mixtral-8x7B-Instruct-v0.1,morning,3.
|
52 |
-
Mixtral-8x7B-Instruct-v0.1,afternoon,5.362903979589355,
|
53 |
-
Mixtral-8x7B-Instruct-v0.1,late afternoon,5.
|
54 |
-
Mixtral-8x7B-Instruct-v0.1,evening,3.6435119574237023,1.
|
55 |
-
Mixtral-8x7B-Instruct-v0.1,late evening,
|
56 |
-
Mixtral-8x7B-Instruct-v0.1,midnight,4.
|
57 |
-
Mixtral-8x7B-Instruct-v0.1,night,4.009439338194697,1.
|
58 |
-
OpenHermes-2.5-Mistral (7B),early morning,2.8495449788040585,
|
59 |
-
OpenHermes-2.5-Mistral (7B),morning,2.
|
60 |
-
OpenHermes-2.5-Mistral (7B),afternoon,3.382803752858152,
|
61 |
-
OpenHermes-2.5-Mistral (7B),late afternoon,2.52796063820521,
|
62 |
-
OpenHermes-2.5-Mistral (7B),evening,2.7269538966092197,
|
63 |
-
OpenHermes-2.5-Mistral (7B),late evening,2.720560759305954,
|
64 |
-
OpenHermes-2.5-Mistral (7B),midnight,2.8086547496470997,
|
65 |
-
OpenHermes-2.5-Mistral (7B),night,2.7470715910196306,1.
|
66 |
-
Qwen 1.5 Chat (7B),early morning,2.167507814036475,
|
67 |
-
Qwen 1.5 Chat (7B),morning,1.
|
68 |
-
Qwen 1.5 Chat (7B),afternoon,2.539412996504042,
|
69 |
-
Qwen 1.5 Chat (7B),late afternoon,2.0715979735056558,1.
|
70 |
-
Qwen 1.5 Chat (7B),evening,2.325377941131592,1.
|
71 |
-
Qwen 1.5 Chat (7B),late evening,1.9339114997697913,1.
|
72 |
-
Qwen 1.5 Chat (7B),midnight,2.2824907505765872,1.
|
73 |
-
Qwen 1.5 Chat (7B),night,1.98169333812518,
|
74 |
-
RedPajama-INCITE Chat (7B),early morning,2.149312309920788,
|
75 |
-
RedPajama-INCITE Chat (7B),morning,1.
|
76 |
-
RedPajama-INCITE Chat (7B),afternoon,1.8110081959854474,
|
77 |
-
RedPajama-INCITE Chat (7B),late afternoon,1.7436921298503876,
|
78 |
-
RedPajama-INCITE Chat (7B),evening,1.964497913013805,
|
79 |
-
RedPajama-INCITE Chat (7B),late evening,2.258105857031686,
|
80 |
-
RedPajama-INCITE Chat (7B),midnight,2.072928147845798,
|
81 |
-
RedPajama-INCITE Chat (7B),night,2.0220184448437815,
|
82 |
-
Snorkel Mistral PairRM DPO (7B),early morning,
|
83 |
-
Snorkel Mistral PairRM DPO (7B),morning,4.
|
84 |
-
Snorkel Mistral PairRM DPO (7B),afternoon,
|
85 |
-
Snorkel Mistral PairRM DPO (7B),late afternoon,
|
86 |
-
Snorkel Mistral PairRM DPO (7B),evening,
|
87 |
-
Snorkel Mistral PairRM DPO (7B),late evening,
|
88 |
-
Snorkel Mistral PairRM DPO (7B),midnight,
|
89 |
-
Snorkel Mistral PairRM DPO (7B),night,
|
90 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,early morning,2.0273348593711855,1.
|
91 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,morning,1.
|
92 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,afternoon,1.8381905496120452,
|
93 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,late afternoon,1.7547113946505954,
|
94 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,evening,1.7984187936782836,
|
95 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,late evening,1.5875422928068372,0.
|
96 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,midnight,1.6341248273849487,0.
|
97 |
-
TinyLlama/TinyLlama-1.1B-Chat-v1.0,night,2.0128010153770446,1.
|
98 |
-
Vicuna v1.5 (7B),early morning,3.3891975168596233,1.
|
99 |
-
Vicuna v1.5 (7B),morning,
|
100 |
-
Vicuna v1.5 (7B),afternoon,3.670576659115878,
|
101 |
-
Vicuna v1.5 (7B),late afternoon,3.768600355495106,1.
|
102 |
-
Vicuna v1.5 (7B),evening,3.0816855430603027,1.
|
103 |
-
Vicuna v1.5 (7B),late evening,3.779375910758972,1.
|
104 |
-
Vicuna v1.5 (7B),midnight,3.8164960656847273,1.
|
105 |
-
Vicuna v1.5 (7B),night,3.648655335108439,1.
|
106 |
-
WizardLM v1.2 (13B),early morning,5.290586443914884,
|
107 |
-
WizardLM v1.2 (13B),morning,4.
|
108 |
-
WizardLM v1.2 (13B),afternoon,7.0199061699127885,
|
109 |
-
WizardLM v1.2 (13B),late afternoon,4.935352149217025,2.
|
110 |
-
WizardLM v1.2 (13B),evening,4.939989816058766,2.
|
111 |
-
WizardLM v1.2 (13B),late evening,5.11581133660816,
|
112 |
-
WizardLM v1.2 (13B),midnight,5.255108930847862,
|
113 |
-
WizardLM v1.2 (13B),night,5.411002513689873,
|
114 |
-
chat-bison (PaLM 2),early morning,2.4665334616388592,2.
|
115 |
-
chat-bison (PaLM 2),morning,2.488477897644043,
|
116 |
-
chat-bison (PaLM 2),afternoon,2.6460144804074215,
|
117 |
-
chat-bison (PaLM 2),late afternoon,3.0387172081253744,2.
|
118 |
-
chat-bison (PaLM 2),evening,2.688272579908371,2.
|
119 |
-
chat-bison (PaLM 2),late evening,2.7250528037548065,
|
120 |
-
chat-bison (PaLM 2),midnight,2.468383938074112,2.
|
121 |
-
chat-bison (PaLM 2),night,2.460119960308075,2.
|
122 |
-
chat-bison-32k (PaLM 2 32K),early morning,9.141417106560299,2.
|
123 |
-
chat-bison-32k (PaLM 2 32K),morning,7.7035503840446475,2.
|
124 |
-
chat-bison-32k (PaLM 2 32K),afternoon,
|
125 |
-
chat-bison-32k (PaLM 2 32K),late afternoon,4.5047362284226855,2.
|
126 |
-
chat-bison-32k (PaLM 2 32K),evening,7.7332194912433625,
|
127 |
-
chat-bison-32k (PaLM 2 32K),late evening,7.
|
128 |
-
chat-bison-32k (PaLM 2 32K),midnight,7.7498266498247785,2.
|
129 |
-
chat-bison-32k (PaLM 2 32K),night,6.491292915344238,2.
|
130 |
-
gemini-pro,early morning,2.7453590130460435,
|
131 |
-
gemini-pro,morning,2.
|
132 |
-
gemini-pro,afternoon,2.816922114008949,
|
133 |
-
gemini-pro,late afternoon,2.9268629640903114,
|
134 |
-
gemini-pro,evening,2.86901999375759,
|
135 |
-
gemini-pro,late evening,3.7189874940246117,
|
136 |
-
gemini-pro,midnight,3.338477972348531,
|
137 |
-
gemini-pro,night,2.8375814715210272,
|
138 |
-
gpt-3.5-turbo,early morning,3.884872453212738,1.
|
139 |
-
gpt-3.5-turbo,morning,3.
|
140 |
-
gpt-3.5-turbo,afternoon,3.9458200880459375,
|
141 |
-
gpt-3.5-turbo,late afternoon,4.384064777692159,1.
|
142 |
-
gpt-3.5-turbo,evening,3.5153889304115657,
|
143 |
-
gpt-3.5-turbo,late evening,
|
144 |
-
gpt-3.5-turbo,midnight,3.822115447632102,
|
145 |
-
gpt-3.5-turbo,night,
|
146 |
-
gpt-4,early morning,14.044063893556595,
|
147 |
-
gpt-4,morning,12.
|
148 |
-
gpt-4,afternoon,16.002364798386893,
|
149 |
-
gpt-4,late afternoon,16.80607506932254,
|
150 |
-
gpt-4,evening,13.841120740345547,
|
151 |
-
gpt-4,late evening,14.
|
152 |
-
gpt-4,midnight,12.3578163115329,
|
153 |
-
gpt-4,night,12.813134506115546,
|
154 |
-
gpt-4-turbo,early morning,
|
155 |
-
gpt-4-turbo,morning,
|
156 |
-
gpt-4-turbo,afternoon,
|
157 |
-
gpt-4-turbo,late afternoon,
|
158 |
-
gpt-4-turbo,evening,
|
159 |
-
gpt-4-turbo,late evening,
|
160 |
-
gpt-4-turbo,midnight,
|
161 |
-
gpt-4-turbo,night,
|
162 |
-
llama-2-70b-chat,early morning,2.8660141522424265,
|
163 |
-
llama-2-70b-chat,morning,2.
|
164 |
-
llama-2-70b-chat,afternoon,4.234376892130426,
|
165 |
-
llama-2-70b-chat,late afternoon,3.6833307104881365,
|
166 |
-
llama-2-70b-chat,evening,2.9706250462084185,
|
167 |
-
llama-2-70b-chat,late evening,4.
|
168 |
-
llama-2-70b-chat,midnight,3.249819871626402,
|
169 |
-
llama-2-70b-chat,night,2.8264514451677147,
|
170 |
-
zephyr-7b-beta,early morning,3.937663261095683,
|
171 |
-
zephyr-7b-beta,morning,4.056525647640228,
|
172 |
-
zephyr-7b-beta,afternoon,3.5789777278900146,
|
173 |
-
zephyr-7b-beta,late afternoon,3.4592524923459447,
|
174 |
-
zephyr-7b-beta,evening,3.5673056403795878,
|
175 |
-
zephyr-7b-beta,late evening,
|
176 |
-
zephyr-7b-beta,midnight,3.7253047794103624,
|
177 |
-
zephyr-7b-beta,night,3.6954557319482166,
|
|
|
1 |
+
model,time_of_day,execution_time_mean,execution_time_0.2_quantile,execution_time_0.4_quantile,execution_time_0.6_quantile,execution_time_0.8_quantile,characters_count_mean,characters_count_0.2_quantile,characters_count_0.4_quantile,characters_count_0.6_quantile,characters_count_0.8_quantile,words_count_mean,words_count_0.2_quantile,words_count_0.4_quantile,words_count_0.6_quantile,words_count_0.8_quantile
|
2 |
+
01-ai Yi Chat (34B),early morning,9.13007491738049,3.884891891479492,4.6730134963989265,11.277822160720826,14.992161560058594,426.6268656716418,339.0,425.0,470.0,483.0,43.656716417910445,15.200000000000001,16.0,59.0,82.0
|
3 |
+
01-ai Yi Chat (34B),morning,11.170560583235725,4.865606260299683,7.8178390026092535,12.402377557754516,16.190500020980835,641.7142857142857,368.6,471.0,564.8,930.4000000000001,82.82539682539682,37.60000000000001,60.0,82.0,122.40000000000002
|
4 |
+
01-ai Yi Chat (34B),afternoon,12.255323665086614,3.8716049194335938,5.125375509262085,9.857397079467773,19.629956483840942,571.9418604651163,335.0,444.0,527.0,717.0,57.593023255813954,15.0,33.0,59.0,84.0
|
5 |
+
01-ai Yi Chat (34B),late afternoon,9.417656523840767,3.9277758598327637,4.811002254486084,13.016066789627075,15.474872350692749,439.95238095238096,339.0,425.0,481.0,483.0,45.857142857142854,16.0,16.0,59.0,82.0
|
6 |
+
01-ai Yi Chat (34B),evening,8.16779062861488,3.7767345905303955,4.79024338722229,9.425769329071045,13.293262481689453,412.3809523809524,297.0,425.0,471.0,481.0,45.0,16.0,16.0,59.0,82.0
|
7 |
+
01-ai Yi Chat (34B),late evening,9.680033167203268,4.015563917160034,5.033138418197632,11.343994951248169,15.70482234954834,424.5,328.6,425.0,471.0,483.0,40.875,15.0,16.0,51.0,82.0
|
8 |
+
01-ai Yi Chat (34B),midnight,9.311747227774726,4.052314758300781,4.877615356445313,10.888091945648194,15.254117584228515,423.53333333333336,338.2,425.0,471.0,483.0,43.111111111111114,15.0,16.0,59.0,82.0
|
9 |
+
01-ai Yi Chat (34B),night,9.625520552260966,3.8614964962005613,5.0270007133483885,11.518787574768066,15.092712736129762,426.9493670886076,339.0,425.0,470.8,485.0,41.9746835443038,15.0,16.0,59.0,80.80000000000001
|
10 |
+
Chronos Hermes (13B),early morning,8.063542557544396,2.4645144939422607,2.6510868072509766,2.99411678314209,15.657413005828857,549.4426229508197,322.0,358.0,375.0,609.0,59.114754098360656,20.0,55.0,55.0,66.0
|
11 |
+
Chronos Hermes (13B),morning,6.32047438621521,2.6101768970489503,3.704946565628052,4.96811203956604,8.575554895401002,586.09375,323.8,412.20000000000005,543.4,746.0,75.0625,44.60000000000001,59.2,67.8,92.80000000000004
|
12 |
+
Chronos Hermes (13B),afternoon,6.651120054020601,2.548703145980835,3.27475643157959,5.613623809814453,9.93179965019226,609.3176470588236,322.0,389.8,544.4,745.0,85.95294117647059,30.600000000000005,56.2,75.19999999999999,105.4
|
13 |
+
Chronos Hermes (13B),late afternoon,7.630303209478205,2.3594475746154786,2.6113294124603272,2.7819708824157714,10.656519556045534,594.1363636363636,322.0,338.2,375.0,1177.0,68.54545454545455,25.0,55.0,59.8,66.0
|
14 |
+
Chronos Hermes (13B),evening,6.468730966250102,2.5155254364013673,2.6828935623168944,2.8508810520172116,8.361930227279663,603.3888888888889,325.0,358.0,375.0,949.8000000000009,80.16666666666667,55.0,55.0,63.0,66.0
|
15 |
+
Chronos Hermes (13B),late evening,7.710561646355523,2.572637176513672,2.74330358505249,2.978392314910889,12.687815237045296,517.1666666666666,323.2,351.40000000000003,375.0,609.0,56.611111111111114,22.0,55.0,56.599999999999994,64.80000000000001
|
16 |
+
Chronos Hermes (13B),midnight,7.710010987520218,2.3319817066192625,2.5218401908874513,2.825849103927612,11.518855381011978,559.5,322.0,344.80000000000007,375.0,1177.0,63.175,24.000000000000004,55.0,58.19999999999999,66.0
|
17 |
+
Chronos Hermes (13B),night,16.544231017430622,2.443719482421875,2.8155813694000242,3.634097862243652,16.203236627578736,586.28,322.0,348.8,375.0,1177.0,68.62666666666667,25.0,55.0,63.0,66.0
|
18 |
+
Falcon Instruct (7B),early morning,1.2471453740046574,0.43880372047424315,0.6281569957733154,0.8937359809875486,1.824258041381836,51.276923076923076,3.0,3.0,83.0,86.0,9.492307692307692,1.0,1.0,14.0,17.0
|
19 |
+
Falcon Instruct (7B),morning,1.4012710821060907,0.44109678268432617,0.6249697208404541,1.7307429313659668,1.833695650100708,54.04761904761905,3.0,73.0,83.0,86.0,10.0,1.0,14.0,14.0,17.0
|
20 |
+
Falcon Instruct (7B),afternoon,1.5774729509611387,0.49752764701843266,0.7844372272491456,1.694784927368164,1.9687790870666506,47.95945945945946,3.0,33.20000000000002,77.0,84.0,9.108108108108109,1.0,6.0,14.0,16.0
|
21 |
+
Falcon Instruct (7B),late afternoon,1.0884015900748116,0.44518136978149414,0.5059421062469482,0.6950340270996094,1.7596561908721924,50.095238095238095,3.0,3.0,83.0,84.0,9.333333333333334,1.0,1.0,14.0,17.0
|
22 |
+
Falcon Instruct (7B),evening,0.7874892950057983,0.43637809753417967,0.47094168663024905,0.6579598426818848,0.7088933944702148,43.55555555555556,3.0,3.0,73.0,83.6,8.222222222222221,1.0,1.0,14.0,16.200000000000003
|
23 |
+
Falcon Instruct (7B),late evening,1.2339241071180864,0.4471440315246582,0.5254935741424561,1.5033053874969482,1.8500621318817139,51.04545454545455,3.0,31.000000000000025,83.0,85.6,9.5,1.0,6.200000000000005,14.0,16.6
|
24 |
+
Falcon Instruct (7B),midnight,0.9382020387894068,0.440291166305542,0.6245289802551269,0.697472095489502,1.158624267578125,47.43589743589744,3.0,3.0,73.0,84.0,8.948717948717949,1.0,1.0,14.0,17.0
|
25 |
+
Falcon Instruct (7B),night,1.2997726936862892,0.4490640163421631,0.6728991985321046,1.1550201416015622,1.866230630874634,51.56164383561644,3.0,3.0,83.0,86.0,9.575342465753424,1.0,1.0,14.0,17.0
|
26 |
+
LLaMA-2 Chat (7B),early morning,2.8209988124230327,2.186510753631592,2.31485333442688,2.902680778503418,3.3846433639526365,470.8970588235294,326.0,382.0,390.0,576.0,64.42647058823529,11.0,59.0,61.0,99.0
|
27 |
+
LLaMA-2 Chat (7B),morning,2.3254783834729875,1.7513995170593262,2.1321896076202393,2.3646205425262448,2.8455467700958255,506.7142857142857,363.8,416.40000000000003,509.59999999999997,621.2,70.88888888888889,52.6,61.0,73.19999999999999,97.0
|
28 |
+
LLaMA-2 Chat (7B),afternoon,3.5475496424569024,2.2166762828826903,2.4947509288787844,3.18677978515625,4.409502410888672,509.81111111111113,325.0,387.6,512.8,636.2000000000002,56.67777777777778,11.8,39.00000000000002,66.0,93.80000000000001
|
29 |
+
LLaMA-2 Chat (7B),late afternoon,2.7755608558654785,2.1486106395721434,2.3446810245513916,3.1260400295257567,3.501351308822632,506.4117647058824,326.0,350.20000000000005,436.0,576.0,73.11764705882354,56.0,59.0,64.6,99.0
|
30 |
+
LLaMA-2 Chat (7B),evening,2.9608939344232734,2.1979090213775634,2.286650037765503,3.350849914550781,4.046070337295532,471.59090909090907,326.0,350.20000000000005,390.0,576.0,68.0,56.0,59.0,64.6,99.0
|
31 |
+
LLaMA-2 Chat (7B),late evening,2.869330001913983,2.20103497505188,2.3568904876708983,2.9004031658172607,3.4293576717376713,472.39130434782606,326.0,371.40000000000003,399.2,576.0,65.82608695652173,29.000000000000014,59.0,62.199999999999996,99.0
|
32 |
+
LLaMA-2 Chat (7B),midnight,2.8870244533457656,2.1818199157714844,2.3481732845306396,3.0668535232543945,3.816838264465333,470.468085106383,326.0,382.0,390.0,576.0,64.51063829787235,11.0,59.0,61.0,99.0
|
33 |
+
LLaMA-2 Chat (7B),night,4.6455250724417265,2.110830783843994,2.762953281402588,3.8129801750183105,5.8875720500946045,453.88524590163934,326.0,382.0,390.0,576.0,62.39344262295082,11.0,59.0,61.0,99.0
|
34 |
+
Mistral (7B) Instruct v0.2 (Together AI),early morning,3.627201789248306,1.9568378925323486,2.339451122283936,3.372861623764038,4.58324556350708,547.0530973451328,294.0,345.0,416.0,514.0,69.01769911504425,6.0,50.0,59.0,73.0
|
35 |
+
Mistral (7B) Instruct v0.2 (Together AI),morning,3.4576284369906864,1.9805867671966553,2.479144334793091,3.364946126937866,4.417264699935913,562.1981981981982,313.0,394.0,469.0,671.0,74.25225225225225,45.0,59.0,68.0,92.0
|
36 |
+
Mistral (7B) Instruct v0.2 (Together AI),afternoon,3.255509059895914,1.170797872543335,1.7924258708953857,2.6285008907318117,4.487909364700319,514.8994708994709,294.0,345.0,449.0,547.0,64.97883597883597,39.0,50.0,67.6,74.0
|
37 |
+
Mistral (7B) Instruct v0.2 (Together AI),late afternoon,3.542298033617545,1.7197493076324464,2.296134614944458,3.5079473495483393,4.537663221359255,567.1449275362319,301.20000000000005,345.0,421.0,514.0,68.5072463768116,7.200000000000003,48.400000000000006,60.0,73.0
|
38 |
+
Mistral (7B) Instruct v0.2 (Together AI),evening,5.496347131400273,2.0290706157684326,2.284267807006836,3.367908000946045,6.334114122390747,542.4827586206897,298.8,345.0,417.0,514.0,68.17241379310344,24.600000000000012,50.0,60.0,73.0
|
39 |
+
Mistral (7B) Instruct v0.2 (Together AI),late evening,3.3933188574654713,1.5389206886291504,2.3081408977508544,3.2702637195587156,4.495786094665528,556.1746031746031,313.0,345.0,450.0,514.0,71.73015873015873,45.0,50.0,68.0,73.0
|
40 |
+
Mistral (7B) Instruct v0.2 (Together AI),midnight,4.510189438914205,2.221144199371338,2.746429681777954,3.95231556892395,5.9356536865234375,545.7692307692307,294.0,345.0,416.0,514.0,68.74725274725274,6.0,50.0,59.0,73.0
|
41 |
+
Mistral (7B) Instruct v0.2 (Together AI),night,3.5451678467578573,1.7408261299133303,2.261455821990967,3.3799757957458496,4.892568922042848,533.1065573770492,294.0,345.0,416.0,514.0,66.94262295081967,6.0,50.0,59.0,73.0
|
42 |
+
Mistral-7B-Instruct-v0.2,early morning,4.625752210617065,4.575008821487427,4.661143350601196,4.697483682632447,4.74144606590271,214.0,97.0,99.0,282.0,339.0,29.75,3.0,5.0,48.0,59.0
|
43 |
+
Mistral-7B-Instruct-v0.2,morning,3.734026002883911,3.334194850921631,3.6873255729675294,4.045523738861084,4.105937862396241,362.9,294.4,337.6,395.6,431.0,58.35,48.8,53.6,60.4,70.0
|
44 |
+
Mistral-7B-Instruct-v0.2,afternoon,3.228973722457886,0.8020155429840088,3.8286522388458253,4.1212786674499515,4.287619972229004,222.0625,0.0,102.0,304.4,421.2,32.25,0.0,3.0,50.8,66.2
|
45 |
+
Mistral-7B-Instruct-v0.2,late afternoon,3.2048643112182615,0.838294506072998,3.8415285110473634,4.12097692489624,4.237715482711792,219.8625,0.0,101.6,304.4,421.2,31.8,0.0,3.0,50.8,66.2
|
46 |
+
Mistral-7B-Instruct-v0.2,evening,3.397640073299408,3.0089535236358644,3.560943269729614,4.059507369995117,4.1680100440979,261.18333333333334,97.0,275.20000000000005,333.4,410.00000000000006,40.1,3.0,46.400000000000006,53.4,68.0
|
47 |
+
Mistral-7B-Instruct-v0.2,late evening,3.389284573495388,0.854732084274292,2.315276288986208,4.109760761260986,4.2783270359039305,175.79375,0.0,51.60000000000012,262.0,398.0,25.68125,0.0,1.8000000000000043,42.0,62.2
|
48 |
+
Mistral-7B-Instruct-v0.2,midnight,2.6972494465964183,0.8847384929656983,0.9363711833953857,4.271179580688477,4.602821731567383,88.14285714285714,0.0,0.0,97.2,102.0,10.607142857142858,0.0,0.0,3.0,5.0
|
49 |
+
Mistral-7B-Instruct-v0.2,night,4.609350740909576,4.585418796539306,4.647099351882934,4.686774873733521,4.740107870101928,214.0,97.0,99.0,282.0,339.0,29.75,3.0,5.0,48.0,59.0
|
50 |
+
Mixtral-8x7B-Instruct-v0.1,early morning,4.526968242530536,1.8919459342956544,2.3256708145141602,3.0963687896728516,5.805924749374391,428.6390977443609,300.0,386.0,434.8,468.0,62.94736842105263,43.0,57.0,66.0,74.0
|
51 |
+
Mixtral-8x7B-Instruct-v0.1,morning,3.786775301242697,1.9155490398406982,2.4697306156158447,3.0977609157562256,4.808599472045898,457.4655172413793,303.0,408.0,457.0,537.0,67.56896551724138,48.0,59.0,69.0,78.0
|
52 |
+
Mixtral-8x7B-Instruct-v0.1,afternoon,5.362903979589355,2.0265902519226073,2.7234819412231444,5.094383001327515,8.19847416877747,452.938679245283,300.0,388.80000000000007,457.0,546.0,66.66037735849056,43.0,58.20000000000002,69.0,88.80000000000001
|
53 |
+
Mixtral-8x7B-Instruct-v0.1,late afternoon,5.80184749175942,2.1974987030029296,3.0992554187774655,5.215971851348877,7.946981716156006,418.3224043715847,298.0,380.79999999999995,433.0,490.0,56.830601092896174,33.99999999999999,48.0,63.400000000000006,77.0
|
54 |
+
Mixtral-8x7B-Instruct-v0.1,evening,3.6435119574237023,1.9721095561981201,2.4149227619171145,3.1797213077545163,4.92767472267151,416.1656050955414,298.0,385.4,433.0,487.0,61.84076433121019,41.0,55.20000000000002,66.0,76.40000000000003
|
55 |
+
Mixtral-8x7B-Instruct-v0.1,late evening,5.62397656769588,1.9343807697296143,2.408257007598877,4.018872165679931,7.493381118774417,476.91379310344826,300.0,376.6000000000001,448.4,589.0,59.172413793103445,7.200000000000003,46.0,62.0,84.80000000000007
|
56 |
+
Mixtral-8x7B-Instruct-v0.1,midnight,4.639010797279158,1.9257502079010012,2.5336658000946044,3.8406158924102782,6.568124961853028,459.7565543071161,298.0,376.80000000000007,457.0,612.0,60.760299625468164,15.0,48.0,66.0,96.60000000000002
|
57 |
+
Mixtral-8x7B-Instruct-v0.1,night,4.009439338194697,1.9190558910369873,2.3876732349395753,3.3867024421691894,6.5049797534942675,436.0131578947368,298.0,375.2000000000001,438.4,492.0,61.10526315789474,33.000000000000014,54.0,66.0,79.60000000000002
|
58 |
+
OpenHermes-2.5-Mistral (7B),early morning,2.8495449788040585,2.0497443675994873,2.3496197700500487,2.9056718826293944,3.602103996276857,281.0416666666667,144.0,287.0,343.0,371.0,32.125,4.0,7.0,57.0,62.0
|
59 |
+
OpenHermes-2.5-Mistral (7B),morning,2.5861338563263416,1.9361390113830566,2.341088151931763,2.7727932929992676,3.219662952423097,385.78125,280.6,344.0,376.4,487.6,51.40625,27.000000000000014,54.400000000000006,58.8,66.20000000000002
|
60 |
+
OpenHermes-2.5-Mistral (7B),afternoon,3.382803752858152,1.9604182720184327,2.2988375663757323,2.848426008224487,3.8551052570343027,369.35483870967744,159.00000000000006,286.8,349.2,475.8,33.8494623655914,4.0,6.6000000000000085,49.0,63.60000000000001
|
61 |
+
OpenHermes-2.5-Mistral (7B),late afternoon,2.52796063820521,1.68584885597229,2.1943514347076416,2.7274121284484862,3.434289741516114,281.125,144.0,287.0,343.0,371.0,32.125,4.0,7.0,57.0,62.0
|
62 |
+
OpenHermes-2.5-Mistral (7B),evening,2.7269538966092197,2.242070960998535,2.4532580852508543,2.736237621307373,3.264935827255249,277.1818181818182,144.0,287.0,343.0,348.0,29.363636363636363,4.0,5.200000000000001,57.0,58.0
|
63 |
+
OpenHermes-2.5-Mistral (7B),late evening,2.720560759305954,2.250868368148804,2.5327273845672607,2.8688658714294433,3.159437894821168,281.125,144.0,287.0,343.0,371.0,32.125,4.0,7.0,57.0,62.0
|
64 |
+
OpenHermes-2.5-Mistral (7B),midnight,2.8086547496470997,2.0280110836029053,2.6033748626708983,2.958510684967041,3.4229323863983154,279.27659574468083,144.0,287.0,343.0,371.0,31.48936170212766,4.0,7.0,57.0,62.0
|
65 |
+
OpenHermes-2.5-Mistral (7B),night,2.7470715910196306,1.8118998527526855,2.4643975257873536,2.7990838527679442,3.451723146438599,281.275,144.0,287.0,343.0,371.0,32.1,4.0,7.0,57.0,62.0
|
66 |
+
Qwen 1.5 Chat (7B),early morning,2.167507814036475,1.3716964721679688,1.9782586574554444,2.1421773433685303,2.5200764179229735,269.0416666666667,183.0,236.0,296.0,368.0,30.125,1.0,4.0,49.0,62.0
|
67 |
+
Qwen 1.5 Chat (7B),morning,2.2512709980919245,1.4333404064178468,1.9432880401611328,2.1577938556671143,2.521956014633179,424.77777777777777,250.4,309.6,397.59999999999997,496.0,55.73015873015873,27.400000000000006,49.0,57.39999999999999,67.6
|
68 |
+
Qwen 1.5 Chat (7B),afternoon,2.539412996504042,1.8534220695495607,2.05458722114563,2.2894588470458985,3.316272544860841,360.26666666666665,197.6,299.0,342.4,442.6,39.03333333333333,3.0,4.0,52.0,67.0
|
69 |
+
Qwen 1.5 Chat (7B),late afternoon,2.0715979735056558,1.4847597122192382,1.9961623191833497,2.1920745372772217,2.350381946563721,269.0,183.0,236.0,296.0,368.0,30.125,1.0,4.0,49.0,62.0
|
70 |
+
Qwen 1.5 Chat (7B),evening,2.325377941131592,1.5857502937316896,2.07436785697937,2.1726004600524904,2.583091640472412,266.59090909090907,183.0,236.0,296.0,355.40000000000003,29.727272727272727,1.6000000000000005,4.0,49.0,59.400000000000006
|
71 |
+
Qwen 1.5 Chat (7B),late evening,1.9339114997697913,1.3509828567504882,1.4851758480072021,2.1028024196624755,2.3062386512756348,264.69565217391306,183.0,226.20000000000005,296.0,342.80000000000007,28.73913043478261,1.0,4.0,49.0,58.40000000000001
|
72 |
+
Qwen 1.5 Chat (7B),midnight,2.2824907505765872,1.469894313812256,2.006262016296387,2.2383018493652345,2.56543288230896,267.72340425531917,183.0,236.0,296.0,368.0,29.638297872340427,1.0,4.0,49.0,62.0
|
73 |
+
Qwen 1.5 Chat (7B),night,1.98169333812518,1.2986116886138916,1.5336131572723388,2.126609468460083,2.4843666553497314,269.2692307692308,183.0,236.0,296.0,368.0,30.17948717948718,1.0,4.0,49.0,62.0
|
74 |
+
RedPajama-INCITE Chat (7B),early morning,2.149312309920788,1.0630450248718262,2.0389476299285887,2.158130168914795,2.622494173049927,60.328125,54.0,54.0,54.0,78.0,5.0,4.0,5.0,5.0,5.0
|
75 |
+
RedPajama-INCITE Chat (7B),morning,1.4803865485721164,0.9543989181518555,1.0027932167053222,1.5951005458831768,2.1152745723724364,59.01587301587302,54.0,54.0,54.0,68.0,5.412698412698413,5.0,5.0,5.0,5.0
|
76 |
+
RedPajama-INCITE Chat (7B),afternoon,1.8110081959854474,1.0464420795440674,1.3511402606964111,2.065592861175537,2.2278335571289065,62.96590909090909,54.0,54.0,57.0,78.0,5.784090909090909,4.0,5.0,5.0,7.0
|
77 |
+
RedPajama-INCITE Chat (7B),late afternoon,1.7436921298503876,1.0311846733093262,2.0036316871643067,2.0985609531402587,2.2179235458374023,59.625,54.0,54.0,54.0,78.0,5.0,4.0,5.0,5.0,5.0
|
78 |
+
RedPajama-INCITE Chat (7B),evening,1.964497913013805,1.0379676342010498,2.0008397102355957,2.0600441455841065,2.1418700218200684,60.13636363636363,54.0,54.0,54.0,78.0,5.0,4.0,5.0,5.0,5.0
|
79 |
+
RedPajama-INCITE Chat (7B),late evening,2.258105857031686,1.0331692695617676,2.0418362617492676,2.116027593612671,3.810591220855713,60.42857142857143,54.0,54.0,54.0,78.0,5.0,4.0,5.0,5.0,5.0
|
80 |
+
RedPajama-INCITE Chat (7B),midnight,2.072928147845798,1.0425399780273437,2.0462307929992676,2.123409461975098,2.2389921665191657,60.0,54.0,54.0,54.0,78.0,5.0,4.0,5.0,5.0,5.0
|
81 |
+
RedPajama-INCITE Chat (7B),night,2.0220184448437815,1.0452874660491944,2.007957410812378,2.078788995742798,2.2428071022033693,59.96153846153846,54.0,54.0,54.0,78.0,5.064102564102564,4.0,5.0,5.0,5.0
|
82 |
+
Snorkel Mistral PairRM DPO (7B),early morning,20.892576156343733,2.556242084503174,3.7271701335906986,5.358461332321166,8.898035669326783,4378.842857142857,464.0,845.0,1088.0,1251.0,103.32857142857142,71.0,80.0,105.0,133.0
|
83 |
+
Snorkel Mistral PairRM DPO (7B),morning,14.711059410125017,4.0283173561096195,4.922686767578125,5.645612335205078,7.975491571426393,3375.03125,926.6000000000001,1151.0,1374.1999999999998,1953.2000000000003,202.640625,105.0,159.0,185.6,235.2000000000001
|
84 |
+
Snorkel Mistral PairRM DPO (7B),afternoon,34.586263124148054,2.558345317840576,4.329371929168701,6.282438659667968,15.324156856536867,6091.4111111111115,657.0,888.2,1251.0,2315.0,577.1555555555556,75.4,104.6,138.0,230.80000000000007
|
85 |
+
Snorkel Mistral PairRM DPO (7B),late afternoon,22.353231539328892,2.429702568054199,3.603127813339234,5.212990999221802,8.90522346496582,4714.0,464.0,845.0,1088.0,1251.0,104.375,71.0,80.0,105.0,133.0
|
86 |
+
Snorkel Mistral PairRM DPO (7B),evening,24.748520737602597,2.34138822555542,3.571380853652954,5.162211179733276,8.841630220413208,5253.047619047619,464.0,845.0,1088.0,1251.0,105.52380952380952,71.0,80.0,105.0,133.0
|
87 |
+
Snorkel Mistral PairRM DPO (7B),late evening,24.120955445549704,2.325694036483765,3.240026903152466,5.769892024993896,8.792698669433594,5043.636363636364,464.0,773.6,1091.6,1251.0,97.77272727272727,71.0,77.6,105.0,123.0
|
88 |
+
Snorkel Mistral PairRM DPO (7B),midnight,22.670283652366475,2.323270893096924,3.7170748710632333,5.51384973526001,8.769575691223144,4791.148936170212,464.0,845.0,1088.0,1251.0,102.82978723404256,71.0,80.0,105.0,133.0
|
89 |
+
Snorkel Mistral PairRM DPO (7B),night,21.554972467422484,2.380740785598755,3.338300323486328,5.203608655929566,8.923078727722169,4539.253333333333,464.0,797.4000000000002,1088.0,1251.0,101.45333333333333,71.0,78.4,105.0,133.0
|
90 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,early morning,2.0273348593711855,1.900433397293091,1.9704366207122803,2.0509723663330077,2.1658658027648925,372.13,334.0,350.0,396.0,420.20000000000005,62.53,57.8,61.0,63.4,72.0
|
91 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,morning,1.9030630230903625,1.8348802089691163,1.8834499359130858,1.904430341720581,1.9469316005706787,358.875,310.6,342.0,394.0,415.4,59.5,48.8,59.0,62.4,71.0
|
92 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,afternoon,1.8381905496120452,1.8016767978668213,1.9038204669952392,1.9715147495269776,2.0281569480896,308.795,280.0,335.8,358.4,415.0,51.08,38.0,57.60000000000001,62.0,71.0
|
93 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,late afternoon,1.7547113946505954,1.7235632419586182,1.8873783588409423,1.9246160984039307,1.9911770820617676,285.17857142857144,85.2,334.0,354.4,412.6,46.9,3.8000000000000007,55.6,61.0,71.0
|
94 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,evening,1.7984187936782836,1.772254180908203,1.8798622608184814,1.9122614860534668,1.9810049057006835,313.99,280.0,338.0,359.0,417.0,51.96,38.0,59.0,62.0,72.0
|
95 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,late evening,1.5875422928068372,0.8799705028533936,1.675227117538452,1.9031693458557128,2.0034618377685547,192.78333333333333,0.0,76.20000000000002,318.0,363.0,31.261111111111113,0.0,3.0,54.0,62.0
|
96 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,midnight,1.6341248273849487,0.8688469886779785,1.912255334854126,2.0132920265197756,2.119899129867554,210.2,0.0,87.80000000000001,334.0,395.0,34.2,0.0,3.6000000000000014,57.4,63.2
|
97 |
+
TinyLlama/TinyLlama-1.1B-Chat-v1.0,night,2.0128010153770446,1.8813746452331543,1.9569619655609132,2.042378377914429,2.157229471206665,372.05,334.0,350.0,396.0,420.20000000000005,62.6,57.8,61.0,63.4,72.0
|
98 |
+
Vicuna v1.5 (7B),early morning,3.3891975168596233,1.60922212600708,2.0776685714721683,2.7636775970458984,5.4549774646759035,327.3157894736842,267.0,349.0,355.0,376.0,37.70175438596491,1.0,4.0,62.0,64.0
|
99 |
+
Vicuna v1.5 (7B),morning,3.4012709193759494,1.606896209716797,2.099214839935303,2.5562299728393554,4.044596385955811,388.4920634920635,268.4,349.0,376.0,443.8,50.38095238095238,28.8,48.0,61.0,66.0
|
100 |
+
Vicuna v1.5 (7B),afternoon,3.670576659115878,2.174506139755249,2.6995410442352297,3.5413179397583003,5.232729816436768,385.25,241.8,336.0,381.0,497.20000000000005,47.77272727272727,2.400000000000002,39.40000000000001,63.0,78.60000000000001
|
101 |
+
Vicuna v1.5 (7B),late afternoon,3.768600355495106,1.6033197879791259,2.519843244552612,3.867197561264038,5.926239156723023,324.27272727272725,278.40000000000003,349.0,354.6,376.0,35.77272727272727,4.0,4.0,61.6,64.0
|
102 |
+
Vicuna v1.5 (7B),evening,3.0816855430603027,1.6141672611236573,1.6617256164550782,3.833245277404785,4.436551856994629,329.0,289.8,354.0,355.0,376.0,42.833333333333336,2.200000000000001,61.0,62.0,64.0
|
103 |
+
Vicuna v1.5 (7B),late evening,3.779375910758972,1.6539848804473878,3.149325180053711,4.361717987060547,5.288036727905274,319.1363636363636,267.0,334.0,354.0,376.0,32.86363636363637,1.0,4.0,61.0,64.0
|
104 |
+
Vicuna v1.5 (7B),midnight,3.8164960656847273,1.6210388660430908,2.529152774810791,4.475259590148926,5.888407230377198,329.9047619047619,278.40000000000003,349.0,355.0,376.0,37.214285714285715,1.6000000000000032,4.0,62.0,64.0
|
105 |
+
Vicuna v1.5 (7B),night,3.648655335108439,1.6238388538360595,2.704656982421875,4.319886350631714,5.733144855499268,321.81944444444446,267.0,349.0,354.0,376.0,34.541666666666664,1.0,4.0,61.0,64.0
|
106 |
+
WizardLM v1.2 (13B),early morning,5.290586443914884,3.126491403579712,3.6899030685424807,5.478811454772949,7.6742314338684094,309.17391304347825,202.0,282.0,305.4,381.0,35.17391304347826,5.0,12.0,48.8,55.0
|
107 |
+
WizardLM v1.2 (13B),morning,5.899739087574066,3.329628658294678,3.9913763046264648,4.60016131401062,5.909883880615235,458.44444444444446,265.8,353.6,427.99999999999994,510.6,71.98412698412699,20.600000000000005,51.0,62.39999999999999,69.6
|
108 |
+
WizardLM v1.2 (13B),afternoon,7.0199061699127885,2.5672261238098146,4.129473257064819,5.420607757568359,8.864709854125977,338.5168539325843,119.0,246.2,311.0,514.2,30.03370786516854,2.0,7.0,22.19999999999999,62.400000000000006
|
109 |
+
WizardLM v1.2 (13B),late afternoon,4.935352149217025,2.9690490245819094,3.663859987258911,5.598448467254638,7.145665264129639,265.0869565217391,202.0,263.0,286.0,311.0,29.043478260869566,5.0,10.600000000000005,46.599999999999994,54.0
|
110 |
+
WizardLM v1.2 (13B),evening,4.939989816058766,2.9392529010772708,3.6680859565734862,5.270061779022217,7.265015745162964,262.1818181818182,202.0,263.0,279.6,311.0,28.136363636363637,5.0,7.8000000000000025,46.0,54.0
|
111 |
+
WizardLM v1.2 (13B),late evening,5.11581133660816,3.2692251205444336,3.8308560848236084,5.560052156448364,7.539101839065552,258.57142857142856,202.0,263.0,286.0,311.0,26.285714285714285,5.0,5.0,45.0,54.0
|
112 |
+
WizardLM v1.2 (13B),midnight,5.255108930847862,3.1566482543945313,4.014303731918335,5.88462200164795,7.7009649753570555,283.20454545454544,202.0,263.0,286.0,311.0,30.681818181818183,5.0,8.799999999999997,46.0,54.0
|
113 |
+
WizardLM v1.2 (13B),night,5.411002513689873,3.2144352436065673,3.8306023597717287,5.623324060440063,8.31432375907898,319.5128205128205,219.60000000000002,285.0,309.4,388.20000000000005,36.84615384615385,5.0,12.0,49.0,61.600000000000016
|
114 |
+
chat-bison (PaLM 2),early morning,2.4665334616388592,2.3049097537994383,2.37662878036499,2.4524997234344483,2.5413723945617677,381.9214285714286,271.2,330.8,407.5999999999997,468.40000000000003,60.892857142857146,43.400000000000006,56.0,65.19999999999997,71.80000000000001
|
115 |
+
chat-bison (PaLM 2),morning,2.488477897644043,2.3304410934448243,2.3861887454986572,2.452661418914795,2.571202039718628,381.73,271.2,330.8,407.59999999999997,468.40000000000003,60.88,43.400000000000006,56.0,65.19999999999999,71.80000000000001
|
116 |
+
chat-bison (PaLM 2),afternoon,2.6460144804074215,2.3701786041259765,2.4414886951446535,2.5810309410095216,2.809173345565796,376.28076923076924,258.0,329.0,392.0,482.0,55.965384615384615,25.0,50.0,64.0,75.0
|
117 |
+
chat-bison (PaLM 2),late afternoon,3.0387172081253744,2.3950292110443114,2.5149539947509765,2.7947920322418214,3.3921805858612064,381.1363636363636,260.0,329.0,392.4,467.20000000000005,53.35454545454545,25.0,49.0,64.0,71.0
|
118 |
+
chat-bison (PaLM 2),evening,2.688272579908371,2.3742778301239014,2.447777843475342,2.5543516635894776,2.8649479866027834,367.07,258.0,329.0,392.0,465.0,55.89,39.0,50.0,64.0,71.0
|
119 |
+
chat-bison (PaLM 2),late evening,2.7250528037548065,2.364246368408203,2.4498497009277345,2.609333801269531,3.047693204879761,382.725,259.6,329.0,404.0,482.80000000000007,52.95,25.0,48.0,64.0,71.0
|
120 |
+
chat-bison (PaLM 2),midnight,2.468383938074112,2.3290175914764406,2.387459325790405,2.430591583251953,2.5517006397247313,381.95,271.2,330.8,407.5999999999997,469.20000000000005,60.9,43.400000000000006,56.0,65.19999999999997,71.80000000000001
|
121 |
+
chat-bison (PaLM 2),night,2.460119960308075,2.296553611755371,2.3889822006225585,2.43837251663208,2.540234899520874,381.92,271.2,330.8,407.59999999999997,469.20000000000005,60.94,43.400000000000006,56.0,65.19999999999999,71.80000000000001
|
122 |
+
chat-bison-32k (PaLM 2 32K),early morning,9.141417106560299,2.6841554641723633,2.8603085041046143,3.1312095165252685,6.370184373855592,335.75,244.2,274.2,323.39999999999975,462.40000000000003,53.85,40.6,45.0,57.0,71.4
|
123 |
+
chat-bison-32k (PaLM 2 32K),morning,7.7035503840446475,2.6868295669555664,2.860622835159302,3.1511935710906984,5.278187990188603,335.75,244.2,274.2,323.4,462.40000000000003,53.85,40.6,45.0,57.0,71.4
|
124 |
+
chat-bison-32k (PaLM 2 32K),afternoon,5.014458654477046,2.7426559925079346,2.885803747177124,3.151959848403931,3.615654516220093,339.3692307692308,231.0,277.0,342.0,460.0,49.323076923076925,32.20000000000004,45.0,55.80000000000001,71.0
|
125 |
+
chat-bison-32k (PaLM 2 32K),late afternoon,4.5047362284226855,2.7416682720184324,2.913498449325562,3.177761459350586,4.069248628616333,348.65909090909093,241.0,277.0,344.40000000000003,472.0,48.35454545454545,31.800000000000004,45.0,54.0,71.0
|
126 |
+
chat-bison-32k (PaLM 2 32K),evening,7.7332194912433625,2.7396387577056887,2.947340488433838,3.21867995262146,5.093518543243415,329.46,239.00000000000006,270.0,311.0,460.0,49.5,37.0,45.0,57.0,71.0
|
127 |
+
chat-bison-32k (PaLM 2 32K),late evening,7.796841062307358,2.7586358547210694,2.9897059917449953,3.3219719886779786,6.034687948226929,349.95,239.00000000000006,277.0,348.4,472.0,47.805,25.0,44.0,53.39999999999999,71.0
|
128 |
+
chat-bison-32k (PaLM 2 32K),midnight,7.7498266498247785,2.6849449634552003,2.841596508026123,3.096329116821289,3.754457473754883,335.75,244.2,274.2,323.39999999999975,462.40000000000003,53.85,40.6,45.0,57.0,71.4
|
129 |
+
chat-bison-32k (PaLM 2 32K),night,6.491292915344238,2.688665819168091,2.857192039489746,3.145037078857422,4.080263996124268,335.75,244.2,274.2,323.4,462.40000000000003,53.85,40.6,45.0,57.0,71.4
|
130 |
+
gemini-pro,early morning,2.7453590130460435,2.2898879051208496,2.4930419921875,2.73421893119812,3.2018725395202643,386.7463768115942,292.0,332.8,387.2,486.0,61.84782608695652,48.0,54.0,61.2,76.0
|
131 |
+
gemini-pro,morning,2.552263767664669,2.196598529815674,2.3861966133117676,2.5183777809143066,3.018606185913086,408.1297709923664,290.0,343.0,403.0,495.0,62.54961832061068,47.0,54.0,61.0,73.0
|
132 |
+
gemini-pro,afternoon,2.816922114008949,2.2006062984466555,2.3862349510192873,2.6984471797943117,3.3357447147369386,372.07142857142856,275.2,324.8,382.2,476.20000000000005,54.948412698412696,41.0,51.0,62.0,75.0
|
133 |
+
gemini-pro,late afternoon,2.9268629640903114,2.199513244628906,2.3799704551696776,2.582952356338501,3.5084349155426033,378.3301886792453,269.2,330.0,382.0,481.8,51.990566037735846,35.2,49.0,59.0,71.0
|
134 |
+
gemini-pro,evening,2.86901999375759,2.314052677154541,2.5094053745269775,2.7460020065307615,3.2915334701538086,375.79487179487177,278.8,332.20000000000005,392.2,463.6,56.45641025641026,43.0,53.0,61.0,72.0
|
135 |
+
gemini-pro,late evening,3.7189874940246117,2.372242975234985,2.7115640163421633,3.359907865524292,4.631186199188233,417.7709923664122,275.4,333.4,396.59999999999997,496.0,54.87786259541985,11.200000000000003,48.0,59.599999999999994,74.0
|
136 |
+
gemini-pro,midnight,3.338477972348531,2.3561188220977782,2.54406418800354,2.85723443031311,3.655383348464966,394.48,283.8,329.6,387.2,485.6,59.64,43.8,53.0,61.39999999999999,75.2
|
137 |
+
gemini-pro,night,2.8375814715210272,2.3033328533172606,2.471330404281616,2.8146450996398924,3.259158134460449,382.44897959183675,295.40000000000003,333.40000000000003,385.2,458.6,61.265306122448976,49.0,54.800000000000004,62.0,72.20000000000002
|
138 |
+
gpt-3.5-turbo,early morning,3.884872453212738,1.7305872917175293,2.143742418289185,2.955885171890259,5.1555583000183125,397.935,243.4,314.6,410.79999999999995,513.0,47.315,1.0,43.0,55.0,72.20000000000002
|
139 |
+
gpt-3.5-turbo,morning,3.411667922410098,1.7421918392181397,2.083978366851807,2.5901238918304443,4.452030658721924,387.3636363636364,255.0,314.0,383.0,473.6,51.236363636363635,33.0,47.0,55.400000000000006,73.0
|
140 |
+
gpt-3.5-turbo,afternoon,3.9458200880459375,1.801002597808838,2.2853939533233643,3.098439598083496,5.724417352676393,381.4428571428571,237.0,292.0,394.4,496.4000000000001,43.76190476190476,1.0,35.0,54.0,74.0
|
141 |
+
gpt-3.5-turbo,late afternoon,4.384064777692159,1.8149762630462647,2.3015675067901613,3.1034206867218015,6.358535432815554,396.4357142857143,241.8,306.6,396.4,499.60000000000014,43.28333333333333,1.0,35.0,54.0,73.20000000000005
|
142 |
+
gpt-3.5-turbo,evening,3.5153889304115657,1.6835531234741212,2.0042774677276616,2.5511348724365233,4.963393735885622,385.18095238095236,244.8,297.0,379.79999999999995,482.4000000000001,46.01428571428571,1.0,42.60000000000002,55.0,73.20000000000005
|
143 |
+
gpt-3.5-turbo,late evening,5.110168156187617,1.6053393840789796,2.0763480186462404,3.583834791183472,6.177494716644288,425.40220820189273,238.0,304.0,403.8,519.8000000000001,45.64511041009464,1.0,31.200000000000017,54.0,76.0
|
144 |
+
gpt-3.5-turbo,midnight,3.822115447632102,1.5931308269500732,2.0406482219696045,3.1707966327667236,5.527797698974609,430.02713178294573,247.0,298.0,404.0,533.0,49.43023255813954,1.0,38.0,55.0,77.0
|
145 |
+
gpt-3.5-turbo,night,6.062970260473398,1.5983599185943604,1.9397829532623292,2.8167493820190432,5.2579593658447275,418.6192307692308,244.8,297.6,412.6,519.6,44.926923076923075,1.0,39.0,55.0,76.0
|
146 |
+
gpt-4,early morning,14.044063893556595,7.295636463165283,9.015167808532714,11.752060985565185,20.17264738082886,328.725,225.60000000000002,293.6,348.79999999999995,424.0000000000001,41.66,1.0,45.0,54.39999999999999,65.20000000000002
|
147 |
+
gpt-4,morning,12.841633839149997,6.992764616012574,9.343756675720217,11.854154300689695,16.941224384307862,356.5662100456621,249.20000000000002,319.4,393.0,451.20000000000005,49.29223744292238,34.6,51.0,60.0,69.4
|
148 |
+
gpt-4,afternoon,16.002364798386893,8.038974475860597,10.440681934356691,14.344744396209716,24.78238425254822,318.3095238095238,214.0,281.6,330.0,412.8000000000002,38.77142857142857,1.0,41.0,53.0,65.20000000000005
|
149 |
+
gpt-4,late afternoon,16.80607506932254,7.710341882705689,10.186593818664551,14.449709796905507,27.238966751098634,314.85885167464113,214.0,272.8,343.2,407.0,37.67224880382775,1.0,39.0,54.0,63.0
|
150 |
+
gpt-4,evening,13.841120740345547,7.064423084259033,8.831003713607789,11.685237216949464,21.826131153106697,318.0809523809524,221.8,286.0,333.4,410.20000000000005,40.19285714285714,1.0,43.0,54.39999999999998,65.0
|
151 |
+
gpt-4,late evening,14.298642643005493,6.543574333190918,8.919902849197388,12.943699979782105,23.03813428878784,315.5983935742972,215.0,273.8,330.0,419.0,37.16064257028113,1.0,26.000000000000284,53.0,67.0
|
152 |
+
gpt-4,midnight,12.3578163115329,5.635044145584106,8.056099367141725,11.609123849868773,17.864221715927123,336.36842105263156,224.0,284.4,338.0,435.0,41.63386727688787,1.0,42.0,55.0,68.0
|
153 |
+
gpt-4,night,12.813134506115546,6.1236782550811775,8.166000318527221,11.773015832901002,18.83016529083252,316.93461538461537,221.60000000000002,282.0,337.0,420.20000000000005,37.93076923076923,1.0,40.60000000000001,54.0,66.0
|
154 |
+
gpt-4-turbo,early morning,11.142605865001679,3.9159236907958985,5.781412267684937,7.572056007385253,11.42907643318177,345.05,147.8,281.20000000000005,358.0,440.20000000000005,45.135,1.0,39.0,52.39999999999999,66.0
|
155 |
+
gpt-4-turbo,morning,13.531417746977372,4.869662284851074,6.702779722213745,8.65960807800293,12.018202924728396,383.5863636363636,240.0,314.8,381.4,445.0,54.9,35.0,48.60000000000001,57.0,66.0
|
156 |
+
gpt-4-turbo,afternoon,13.997754749229976,4.34612717628479,6.371346712112427,9.331586408615111,15.885145711898806,351.01190476190476,192.40000000000003,280.6,359.19999999999993,447.0,46.03333333333333,3.8000000000000114,39.60000000000002,53.0,67.20000000000005
|
157 |
+
gpt-4-turbo,late afternoon,22.320911452883767,4.842201137542725,7.02568325996399,9.829535150527954,17.218585824966432,381.65714285714284,206.8,291.0,354.4,454.20000000000005,47.35476190476191,3.0,40.0,53.39999999999998,68.20000000000005
|
158 |
+
gpt-4-turbo,evening,16.550320884159632,4.094121646881104,5.7816534519195555,7.8970251083374015,14.118882179260256,382.31666666666666,206.00000000000006,297.6,363.0,456.0,48.45,3.8000000000000114,42.0,55.0,69.20000000000005
|
159 |
+
gpt-4-turbo,late evening,14.592236209392548,3.699828004837036,4.998922681808472,7.392157745361327,13.814279460906983,413.61,203.60000000000002,291.0,361.0,453.60000000000014,44.8,2.0,36.200000000000045,52.0,66.0
|
160 |
+
gpt-4-turbo,midnight,13.770663784850727,3.6464269638061526,4.955322504043579,7.001602125167845,12.311533260345465,382.7613636363636,200.8,295.20000000000005,377.4,458.4000000000001,47.61818181818182,3.0,39.0,55.0,69.20000000000005
|
161 |
+
gpt-4-turbo,night,14.254795966698573,3.4461984157562258,4.772410726547242,6.590600490570068,12.257953500747684,352.54615384615386,117.80000000000001,287.0,375.0,447.0,46.37692307692308,1.0,39.60000000000001,55.400000000000006,67.0
|
162 |
+
llama-2-70b-chat,early morning,2.8660141522424265,2.252003002166748,2.5588028907775877,2.866885948181152,3.3538597583770753,355.7280701754386,281.0,303.0,349.0,469.0,54.78947368421053,43.0,46.0,51.0,71.0
|
163 |
+
llama-2-70b-chat,morning,2.6787980123728263,1.9940367698669434,2.3450462341308596,2.6430736064910887,3.3015005588531494,386.12605042016804,281.0,310.2,391.79999999999995,480.0,57.403361344537814,43.0,46.0,55.8,73.00000000000003
|
164 |
+
llama-2-70b-chat,afternoon,4.234376892130426,2.2855687141418457,2.7579305171966553,3.8900455951690676,6.170413398742677,415.3872340425532,255.0,303.0,376.0,553.0,61.11489361702127,34.0,46.0,62.0,81.2000000000001
|
165 |
+
llama-2-70b-chat,late afternoon,3.6833307104881365,2.2703349590301514,2.6599215507507323,3.1460047721862794,4.00149941444397,481.8484848484849,267.20000000000016,310.0,376.0,582.0,66.05050505050505,34.0,45.0,53.0,79.60000000000002
|
166 |
+
llama-2-70b-chat,evening,2.9706250462084185,2.3251545429229736,2.602426767349243,2.9736311435699463,3.518934726715088,350.54696132596683,255.0,300.0,338.0,469.0,52.3756906077348,34.0,45.0,51.0,71.0
|
167 |
+
llama-2-70b-chat,late evening,4.719581684340602,2.4600090980529785,2.9465444087982178,3.631521224975586,5.516965627670288,601.1920289855072,281.0,311.0,473.0,707.0,83.8804347826087,34.0,46.0,65.0,102.0
|
168 |
+
llama-2-70b-chat,midnight,3.249819871626402,2.2821081638336183,2.6177178382873536,2.9772785186767576,3.9292076110839846,365.14473684210526,255.0,300.0,349.0,473.0,54.80921052631579,34.0,45.0,53.0,77.0
|
169 |
+
llama-2-70b-chat,night,2.8264514451677147,2.1450907707214357,2.5386868476867677,2.966430997848511,3.449947023391724,356.4886363636364,281.0,303.0,349.0,469.0,55.02272727272727,43.0,46.0,51.39999999999999,71.0
|
170 |
+
zephyr-7b-beta,early morning,3.937663261095683,3.9492252349853514,4.371797418594361,4.507887315750122,4.795662927627563,273.03333333333336,86.8,321.20000000000005,376.6,404.80000000000007,43.45,3.0,51.0,60.0,68.0
|
171 |
+
zephyr-7b-beta,morning,4.056525647640228,4.067356967926026,4.103496074676514,4.111767244338989,4.127487230300903,386.3,354.8,377.4,397.2,434.6,63.35,54.8,61.800000000000004,67.0,71.2
|
172 |
+
zephyr-7b-beta,afternoon,3.5789777278900146,3.701544761657715,4.1053015232086185,4.16576886177063,4.2883219718933105,277.6,86.8,322.0,386.4,422.4,44.016666666666666,3.0,53.6,62.4,68.0
|
173 |
+
zephyr-7b-beta,late afternoon,3.4592524923459447,2.378071928024296,4.1054285049438475,4.143296527862549,4.255803108215332,251.1010101010101,48.600000000000115,147.60000000000048,375.0,406.4,39.292929292929294,0.6000000000000014,27.000000000000085,59.8,68.0
|
174 |
+
zephyr-7b-beta,evening,3.5673056403795878,3.713316535949707,4.0996331691741945,4.1397500991821286,4.223864793777466,273.3666666666667,86.8,321.20000000000005,376.6,404.80000000000007,43.5,3.0,51.0,61.199999999999996,68.0
|
175 |
+
zephyr-7b-beta,late evening,3.45343524068594,0.8169618606567383,2.4506575584411654,4.131957244873047,4.292169427871704,176.23125,0.0,48.600000000000115,180.19999999999862,396.0,27.21875,0.0,0.6000000000000014,32.999999999999744,65.0
|
176 |
+
zephyr-7b-beta,midnight,3.7253047794103624,1.2049674987792969,4.328699159622192,4.623009824752808,4.870654821395874,217.82083333333333,0.0,91.0,359.4,403.0,33.67916666666667,0.0,3.6000000000000085,58.0,67.0
|
177 |
+
zephyr-7b-beta,night,3.6954557319482166,1.013876724243164,4.408831071853638,4.585428762435913,4.776888847351074,216.55833333333334,0.0,98.0,350.0,396.6,33.46666666666667,0.0,3.6000000000000014,54.39999999999999,67.0
|
data/time_of_day_plots.csv
CHANGED
The diff for this file is too large to render.
See raw diff
|
|
pipeline/config.py
CHANGED
@@ -55,8 +55,9 @@ class ModelCostsConfig(Config):
|
|
55 |
|
56 |
|
57 |
class TimeOfDayComparisonConfig(Config):
|
|
|
58 |
saving_path: str = "data/"
|
59 |
-
|
60 |
|
61 |
class GeneralPlotConfig(Config):
|
62 |
plots_dir: str = "./html/plots/"
|
|
|
55 |
|
56 |
|
57 |
class TimeOfDayComparisonConfig(Config):
|
58 |
+
plots_dir: str = "./html/plots/"
|
59 |
saving_path: str = "data/"
|
60 |
+
quantiles: List[float] = [0.2, 0.4, 0.6, 0.8]
|
61 |
|
62 |
class GeneralPlotConfig(Config):
|
63 |
plots_dir: str = "./html/plots/"
|
pipeline/models.py
CHANGED
@@ -107,7 +107,6 @@ MODELS = [
|
|
107 |
"Hugging Face Inference Endpoint",
|
108 |
hourly_cost=1.30,
|
109 |
size_billion_parameters=7,
|
110 |
-
selected=True,
|
111 |
),
|
112 |
Model(
|
113 |
"TinyLlama/TinyLlama-1.1B-Chat-v1.0",
|
@@ -513,16 +512,11 @@ MODELS = [
|
|
513 |
),
|
514 |
]
|
515 |
|
516 |
-
MISSING_MEASUREMENTS = [
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
'01-ai Yi Chat (34B)', 'Chronos Hermes (13B)', 'LLaMA-2 Chat (7B)',
|
522 |
-
'Qwen 1.5 Chat (7B)', 'Snorkel Mistral PairRM DPO (7B)',
|
523 |
-
'OpenHermes-2.5-Mistral (7B)', 'Falcon Instruct (7B)',
|
524 |
-
'RedPajama-INCITE Chat (7B)', 'WizardLM v1.2 (13B)',
|
525 |
-
'Upstage SOLAR Instruct v1 (11B)']
|
526 |
|
527 |
def together_ai_price(billions_of_parameters):
|
528 |
if billions_of_parameters <= 4:
|
|
|
107 |
"Hugging Face Inference Endpoint",
|
108 |
hourly_cost=1.30,
|
109 |
size_billion_parameters=7,
|
|
|
110 |
),
|
111 |
Model(
|
112 |
"TinyLlama/TinyLlama-1.1B-Chat-v1.0",
|
|
|
512 |
),
|
513 |
]
|
514 |
|
515 |
+
MISSING_MEASUREMENTS = [
|
516 |
+
"gemini-pro",
|
517 |
+
"chat-bison (PaLM 2)",
|
518 |
+
"chat-bison-32k (PaLM 2 32K)"
|
519 |
+
]
|
|
|
|
|
|
|
|
|
|
|
520 |
|
521 |
def together_ai_price(billions_of_parameters):
|
522 |
if billions_of_parameters <= 4:
|