File size: 16,816 Bytes
87f4ccf
 
 
 
 
d63951d
 
87f4ccf
 
 
8b82664
d63951d
 
 
87f4ccf
 
 
 
ee5155f
0481eff
 
 
ee5155f
0481eff
 
 
08d3f07
0481eff
08d3f07
 
 
0040773
87f4ccf
 
 
 
0481eff
 
a798e0c
34471a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a798e0c
 
018014a
 
a798e0c
 
 
 
 
 
 
 
 
 
 
 
018014a
 
 
 
a798e0c
 
 
 
 
 
018014a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ee5155f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a798e0c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
018014a
 
 
0481eff
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
---
base_model: upstage/SOLAR-10.7B-Instruct-v1.0
tags:
- alignment-handbook
- generated_from_trainer
- UNA
- single-turn
model-index:
- name: UNA-SOLAR-10.7B-Instruct-v1.0
  results: []
license: cc-by-nc-nd-4.0
language:
- en
library_name: transformers
---

# UNA: Uniform Neural Alignment

SFT Further:
- Linear
- 2e-5

Merges:
- Fan in: `0:2`
- Fan out: `-4:`
- Intermediary layers: `1/1/1/0/1/1/0/1/0/1/1/0/1/1/0` use the On/Off as a way of regularise.
## Quants

* [ggml-model-q5_k_m.gguf](https://huggingface.co/fblgit/UNA-SOLAR-10.7B-Instruct-v1.0/resolve/main/ggml-model-q5_k_m.gguf?download=true)
* [ggml-model-q6_k.gguf](https://huggingface.co/fblgit/UNA-SOLAR-10.7B-Instruct-v1.0/resolve/main/ggml-model-q6_k.gguf?download=true)
  
## Libraries:

- Transformers 4.35.0-UNA
- Pytorch 2.1.0
- Datasets 2.14.6
- Tokenizers 0.14.1

## Evals LM-Evaluation Harness
`mt-bench`:
```
Mode: single
Input file: data/mt_bench/model_judgment/gpt-4_single.jsonl

########## First turn ##########
                                      score
model                         turn
gpt-4                         1     8.95625
claude-v1                     1     8.15000
gpt-3.5-turbo                 1     8.07500
LUNA-SOLARkrautLM-Instruct    1     7.93750
UNA-SOLAR-10.7B-Instruct-v1.0 1     7.80625
vicuna-33b-v1.3               1     7.45625
wizardlm-30b                  1     7.13125
tulu-30b                      1     7.01875
vicuna-13b-v1.3               1     6.81250
guanaco-65b                   1     6.78125
nous-hermes-13b               1     6.43125
alpaca-13b                    1     4.97500
rwkv-4-raven-14b              1     4.74375
llama-13b                     1     3.26250

########## Second turn ##########
                                       score
model                         turn
gpt-4                         2     9.025000
gpt-3.5-turbo                 2     7.812500
claude-v1                     2     7.650000
UNA-SOLAR-10.7B-Instruct-v1.0 2     7.237500
LUNA-SOLARkrautLM-Instruct    2     6.987500
wizardlm-30b                  2     6.887500
vicuna-33b-v1.3               2     6.787500
guanaco-65b                   2     6.037500
vicuna-13b-v1.3               2     5.962500
tulu-30b                      2     5.850000
nous-hermes-13b               2     4.664557
alpaca-13b                    2     4.087500
rwkv-4-raven-14b              2     3.225000
llama-13b                     2     1.950000

########## Average ##########
                                  score
model
gpt-4                          8.990625
gpt-3.5-turbo                  7.943750
claude-instant-v1              7.905660
claude-v1                      7.900000
UNA-SOLAR-10.7B-Instruct-v1.0  7.521875
LUNA-SOLARkrautLM-Instruct     7.462500
vicuna-33b-v1.3                7.121875
wizardlm-30b                   7.009375
Llama-2-70b-chat               6.856250
Llama-2-13b-chat               6.650000
guanaco-33b                    6.528125
tulu-30b                       6.434375
guanaco-65b                    6.409375
oasst-sft-7-llama-30b          6.409375
palm-2-chat-bison-001          6.400000
mpt-30b-chat                   6.393750
vicuna-13b-v1.3                6.387500
wizardlm-13b                   6.353125
Llama-2-7b-chat                6.268750
vicuna-7b-v1.3                 5.996875
baize-v2-13b                   5.750000
nous-hermes-13b                5.553459
mpt-7b-chat                    5.459119
gpt4all-13b-snoozy             5.452830
koala-13b                      5.350000
mpt-30b-instruct               5.218750
falcon-40b-instruct            5.168750
h2ogpt-oasst-open-llama-13b    4.625000
alpaca-13b                     4.531250
chatglm-6b                     4.500000
oasst-sft-4-pythia-12b         4.318750
rwkv-4-raven-14b               3.984375
dolly-v2-12b                   3.275000
fastchat-t5-3b                 3.040625
stablelm-tuned-alpha-7b        2.753125
llama-13b                      2.606250
```

`big-refactor` branch:

```
hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0), gen_kwargs: (None), limit: None, num_fewshot: 25, batch_size: auto (32)
|    Tasks    |Version|Filter|n-shot| Metric |Value |   |Stderr|
|-------------|-------|------|-----:|--------|-----:|---|-----:|
|arc_challenge|Yaml   |none  |    25|acc     |0.6954|±  |0.0134|
|             |       |none  |    25|acc_norm|0.7167|±  |0.0132|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|  Filter  |n-shot|  Metric   |Value|   |Stderr|
|-----|-------|----------|-----:|-----------|----:|---|-----:|
|gsm8k|Yaml   |get-answer|     5|exact_match|0.671|±  |0.0129|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0), gen_kwargs: (), limit: None, num_fewshot: 0, batch_size: auto (64)
|    Tasks     |Version|Filter|n-shot|Metric|Value |   |Stderr|
|--------------|-------|------|-----:|------|-----:|---|-----:|
|truthfulqa_mc2|Yaml   |none  |     0|acc   |0.7297|_  |0.0149|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0), gen_kwargs: (None), limit: None, num_fewshot: 10, batch_size: auto (32)
|  Tasks  |Version|Filter|n-shot| Metric |Value |   |Stderr|
|---------|-------|------|-----:|--------|-----:|---|-----:|
|hellaswag|Yaml   |none  |    10|acc     |0.7091|±  |0.0045|
|         |       |none  |    10|acc_norm|0.8821|±  |0.0032|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0,dtype=float16), gen_kwargs: (), limit: None, num_fewshot: 0, batch_size: auto (32)
|    Tasks     |Version|Filter|n-shot|  Metric  |Value |   |Stderr|
|--------------|-------|------|-----:|----------|-----:|---|-----:|
|boolq         |Yaml   |none  |     0|acc       |0.8807|_  |0.0057|
|lambada_openai|Yaml   |none  |     0|perplexity|3.2452|_  |0.0778|
|              |       |none  |     0|acc       |0.7207|_  |0.0063|
|piqa          |Yaml   |none  |     0|acc       |0.8020|_  |0.0093|
|              |       |none  |     0|acc_norm  |0.8009|_  |0.0093|
|sciq          |Yaml   |none  |     0|acc       |0.9730|_  |0.0051|
|              |       |none  |     0|acc_norm  |0.9630|_  |0.0060|
|winogrande    |Yaml   |none  |     0|acc       |0.7577|_  |0.0120|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0,dtype=float16), gen_kwargs: (), limit: None, num_fewshot: 0, batch_size: auto (64)
| Tasks  |Version|Filter|n-shot| Metric |Value |   |Stderr|
|--------|-------|------|-----:|--------|-----:|---|-----:|
|mathqa  |Yaml   |none  |     0|acc     |0.3474|_  |0.0087|
|        |       |none  |     0|acc_norm|0.3568|_  |0.0088|
|pubmedqa|Yaml   |none  |     0|acc     |0.5400|_  |0.0223|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0,dtype=float16), gen_kwargs: (), limit: None, num_fewshot: 0, batch_size: auto
|                        Tasks                         |Version|Filter|n-shot|  Metric   |Value |   |Stderr|
|------------------------------------------------------|-------|------|-----:|-----------|-----:|---|-----:|
|bbh_fewshot                                           |N/A    |none  |     0|exact_match|0.4660|_  |0.1771|
| - bbh_fewshot_boolean_expressions                    |Yaml   |none  |     0|exact_match|0.8160|_  |0.0246|
| - bbh_fewshot_causal_judgement                       |Yaml   |none  |     0|exact_match|0.4973|_  |0.0367|
| - bbh_fewshot_date_understanding                     |Yaml   |none  |     0|exact_match|0.4840|_  |0.0317|
| - bbh_fewshot_disambiguation_qa                      |Yaml   |none  |     0|exact_match|0.6520|_  |0.0302|
| - bbh_fewshot_dyck_languages                         |Yaml   |none  |     0|exact_match|0.2040|_  |0.0255|
| - bbh_fewshot_formal_fallacies                       |Yaml   |none  |     0|exact_match|0.5280|_  |0.0316|
| - bbh_fewshot_geometric_shapes                       |Yaml   |none  |     0|exact_match|0.3360|_  |0.0299|
| - bbh_fewshot_hyperbaton                             |Yaml   |none  |     0|exact_match|0.5520|_  |0.0315|
| - bbh_fewshot_logical_deduction_five_objects         |Yaml   |none  |     0|exact_match|0.4520|_  |0.0315|
| - bbh_fewshot_logical_deduction_seven_objects        |Yaml   |none  |     0|exact_match|0.3920|_  |0.0309|
| - bbh_fewshot_logical_deduction_three_objects        |Yaml   |none  |     0|exact_match|0.6200|_  |0.0308|
| - bbh_fewshot_movie_recommendation                   |Yaml   |none  |     0|exact_match|0.6640|_  |0.0299|
| - bbh_fewshot_multistep_arithmetic_two               |Yaml   |none  |     0|exact_match|0.0080|_  |0.0056|
| - bbh_fewshot_navigate                               |Yaml   |none  |     0|exact_match|0.6280|_  |0.0306|
| - bbh_fewshot_object_counting                        |Yaml   |none  |     0|exact_match|0.3960|_  |0.0310|
| - bbh_fewshot_penguins_in_a_table                    |Yaml   |none  |     0|exact_match|0.4726|_  |0.0415|
| - bbh_fewshot_reasoning_about_colored_objects        |Yaml   |none  |     0|exact_match|0.5320|_  |0.0316|
| - bbh_fewshot_ruin_names                             |Yaml   |none  |     0|exact_match|0.5680|_  |0.0314|
| - bbh_fewshot_salient_translation_error_detection    |Yaml   |none  |     0|exact_match|0.5480|_  |0.0315|
| - bbh_fewshot_snarks                                 |Yaml   |none  |     0|exact_match|0.5169|_  |0.0376|
| - bbh_fewshot_sports_understanding                   |Yaml   |none  |     0|exact_match|0.8320|_  |0.0237|
| - bbh_fewshot_temporal_sequences                     |Yaml   |none  |     0|exact_match|0.5520|_  |0.0315|
| - bbh_fewshot_tracking_shuffled_objects_five_objects |Yaml   |none  |     0|exact_match|0.1480|_  |0.0225|
| - bbh_fewshot_tracking_shuffled_objects_seven_objects|Yaml   |none  |     0|exact_match|0.1720|_  |0.0239|
| - bbh_fewshot_tracking_shuffled_objects_three_objects|Yaml   |none  |     0|exact_match|0.2760|_  |0.0283|
| - bbh_fewshot_web_of_lies                            |Yaml   |none  |     0|exact_match|0.4760|_  |0.0316|
| - bbh_fewshot_word_sorting                           |Yaml   |none  |     0|exact_match|0.2840|_  |0.0286|

|  Groups   |Version|Filter|n-shot|  Metric   |Value|   |Stderr|
|-----------|-------|------|-----:|-----------|----:|---|-----:|
|bbh_fewshot|N/A    |none  |     0|exact_match|0.466|_  |0.1771|

hf (pretrained=fblgit/UNA-SOLAR-10.7B-Instruct-v1.0), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto (16)
|                 Tasks                 |Version|Filter|n-shot|Metric|Value |   |Stderr|
|---------------------------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu                                   |N/A    |none  |     0|acc   |0.6513|±  |0.1221|
| - humanities                          |N/A    |none  |     5|acc   |0.6077|±  |0.1185|
|  - formal_logic                       |Yaml   |none  |     5|acc   |0.4444|±  |0.0444|
|  - high_school_european_history       |Yaml   |none  |     5|acc   |0.8121|±  |0.0305|
|  - high_school_us_history             |Yaml   |none  |     5|acc   |0.8431|±  |0.0255|
|  - high_school_world_history          |Yaml   |none  |     5|acc   |0.8523|±  |0.0231|
|  - international_law                  |Yaml   |none  |     5|acc   |0.7851|±  |0.0375|
|  - jurisprudence                      |Yaml   |none  |     5|acc   |0.7870|±  |0.0396|
|  - logical_fallacies                  |Yaml   |none  |     5|acc   |0.7546|±  |0.0338|
|  - moral_disputes                     |Yaml   |none  |     5|acc   |0.7370|±  |0.0237|
|  - moral_scenarios                    |Yaml   |none  |     5|acc   |0.4101|±  |0.0164|
|  - philosophy                         |Yaml   |none  |     5|acc   |0.7170|±  |0.0256|
|  - prehistory                         |Yaml   |none  |     5|acc   |0.7840|±  |0.0229|
|  - professional_law                   |Yaml   |none  |     5|acc   |0.4941|±  |0.0128|
|  - world_religions                    |Yaml   |none  |     5|acc   |0.7895|±  |0.0313|
| - other                               |N/A    |none  |     5|acc   |0.7116|±  |0.0939|
|  - business_ethics                    |Yaml   |none  |     5|acc   |0.7600|±  |0.0429|
|  - clinical_knowledge                 |Yaml   |none  |     5|acc   |0.6792|±  |0.0287|
|  - college_medicine                   |Yaml   |none  |     5|acc   |0.6590|±  |0.0361|
|  - global_facts                       |Yaml   |none  |     5|acc   |0.3400|±  |0.0476|
|  - human_aging                        |Yaml   |none  |     5|acc   |0.6816|±  |0.0313|
|  - management                         |Yaml   |none  |     5|acc   |0.8350|±  |0.0368|
|  - marketing                          |Yaml   |none  |     5|acc   |0.8547|±  |0.0231|
|  - medical_genetics                   |Yaml   |none  |     5|acc   |0.7000|±  |0.0461|
|  - miscellaneous                      |Yaml   |none  |     5|acc   |0.8020|±  |0.0142|
|  - nutrition                          |Yaml   |none  |     5|acc   |0.7418|±  |0.0251|
|  - professional_accounting            |Yaml   |none  |     5|acc   |0.5071|±  |0.0298|
|  - professional_medicine              |Yaml   |none  |     5|acc   |0.7500|±  |0.0263|
|  - virology                           |Yaml   |none  |     5|acc   |0.5843|±  |0.0384|
| - social_sciences                     |N/A    |none  |     5|acc   |0.7537|±  |0.0681|
|  - econometrics                       |Yaml   |none  |     5|acc   |0.5000|±  |0.0470|
|  - high_school_geography              |Yaml   |none  |     5|acc   |0.8586|±  |0.0248|
|  - high_school_government_and_politics|Yaml   |none  |     5|acc   |0.9016|±  |0.0215|
|  - high_school_macroeconomics         |Yaml   |none  |     5|acc   |0.6615|±  |0.0240|
|  - high_school_microeconomics         |Yaml   |none  |     5|acc   |0.7311|±  |0.0288|
|  - high_school_psychology             |Yaml   |none  |     5|acc   |0.8404|±  |0.0157|
|  - human_sexuality                    |Yaml   |none  |     5|acc   |0.7328|±  |0.0388|
|  - professional_psychology            |Yaml   |none  |     5|acc   |0.6814|±  |0.0189|
|  - public_relations                   |Yaml   |none  |     5|acc   |0.6909|±  |0.0443|
|  - security_studies                   |Yaml   |none  |     5|acc   |0.7469|±  |0.0278|
|  - sociology                          |Yaml   |none  |     5|acc   |0.8308|±  |0.0265|
|  - us_foreign_policy                  |Yaml   |none  |     5|acc   |0.8900|±  |0.0314|
| - stem                                |N/A    |none  |     5|acc   |0.5569|±  |0.1380|
|  - abstract_algebra                   |Yaml   |none  |     5|acc   |0.4100|±  |0.0494|
|  - anatomy                            |Yaml   |none  |     5|acc   |0.6222|±  |0.0419|
|  - astronomy                          |Yaml   |none  |     5|acc   |0.7368|±  |0.0358|
|  - college_biology                    |Yaml   |none  |     5|acc   |0.8056|±  |0.0331|
|  - college_chemistry                  |Yaml   |none  |     5|acc   |0.4700|±  |0.0502|
|  - college_computer_science           |Yaml   |none  |     5|acc   |0.5100|±  |0.0502|
|  - college_mathematics                |Yaml   |none  |     5|acc   |0.2800|±  |0.0451|
|  - college_physics                    |Yaml   |none  |     5|acc   |0.3431|±  |0.0472|
|  - computer_security                  |Yaml   |none  |     5|acc   |0.7400|±  |0.0441|
|  - conceptual_physics                 |Yaml   |none  |     5|acc   |0.6340|±  |0.0315|
|  - electrical_engineering             |Yaml   |none  |     5|acc   |0.6000|±  |0.0408|
|  - elementary_mathematics             |Yaml   |none  |     5|acc   |0.4815|±  |0.0257|
|  - high_school_biology                |Yaml   |none  |     5|acc   |0.8032|±  |0.0226|
|  - high_school_chemistry              |Yaml   |none  |     5|acc   |0.4877|±  |0.0352|
|  - high_school_computer_science       |Yaml   |none  |     5|acc   |0.7200|±  |0.0451|
|  - high_school_mathematics            |Yaml   |none  |     5|acc   |0.3815|±  |0.0296|
|  - high_school_physics                |Yaml   |none  |     5|acc   |0.3576|±  |0.0391|
|  - high_school_statistics             |Yaml   |none  |     5|acc   |0.5602|±  |0.0339|
|  - machine_learning                   |Yaml   |none  |     5|acc   |0.4643|±  |0.0473|

|      Groups      |Version|Filter|n-shot|Metric|Value |   |Stderr|
|------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu              |N/A    |none  |     0|acc   |0.6513|±  |0.1221|
| - humanities     |N/A    |none  |     5|acc   |0.6077|±  |0.1185|
| - other          |N/A    |none  |     5|acc   |0.7116|±  |0.0939|
| - social_sciences|N/A    |none  |     5|acc   |0.7537|±  |0.0681|
| - stem           |N/A    |none  |     5|acc   |0.5569|±  |0.1380|
```


## Citations

to [Upstage.AI](https://huggingface.co/upstage) for its awesome base model, this is merely a UNA of it. It can only refine what its already in there :)

If you find UNA-SOLAR useful, cite and support the authors.