upadted README and config
Browse files- README.md +101 -0
- config.json +4 -4
README.md
CHANGED
@@ -1,3 +1,104 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
pipeline_tag: text-generation
|
4 |
+
library_name: transformers
|
5 |
+
language: [
|
6 |
+
'am', 'ar', 'as', 'az', 'be', 'bg', 'bn', 'bn_rom', 'br',
|
7 |
+
'bs', 'ca', 'cs', 'cy', 'da', 'de', 'el', 'en', 'eo', 'es',
|
8 |
+
'et', 'eu', 'fa', 'ff', 'fi', 'fr', 'fy', 'ga', 'gd', 'gl',
|
9 |
+
'gn', 'gu', 'ha', 'he', 'hi', 'hi_rom', 'hr', 'ht', 'hu',
|
10 |
+
'hy', 'id', 'ig', 'is', 'it', 'ja', 'jv', 'ka', 'kk', 'km',
|
11 |
+
'kn', 'ko', 'ku', 'ky', 'la', 'lg', 'li', 'ln', 'lo', 'lt',
|
12 |
+
'lv', 'mg', 'mk', 'ml', 'mn', 'mr', 'ms', 'my', 'my_zaw',
|
13 |
+
'ne', 'nl', 'no', 'ns', 'om', 'or', 'pa', 'pl', 'ps', 'pt',
|
14 |
+
'qu', 'rm', 'ro', 'ru', 'sa', 'si', 'sc', 'sd', 'sk', 'sl',
|
15 |
+
'so', 'sq', 'sr', 'ss', 'su', 'sv', 'sw', 'ta', 'ta_rom',
|
16 |
+
'te', 'te_rom', 'th', 'tl', 'tn', 'tr', 'ug', 'uk', 'ur',
|
17 |
+
'ur_rom', 'uz', 'vi', 'wo', 'xh', 'yi', 'yo',
|
18 |
+
'zh-Hans', 'zh-Hant', 'zu',
|
19 |
+
]
|
20 |
+
datasets: [
|
21 |
+
'yahma/alpaca-cleaned',
|
22 |
+
'gbharti/wealth-alpaca_lora',
|
23 |
+
'databricks/databricks-dolly-15k',
|
24 |
+
'VMware/open-instruct',
|
25 |
+
'saillab/taco-datasets',
|
26 |
+
'xu-song/cc100-samples',
|
27 |
+
'jordiclive/wikipedia-summary-dataset',
|
28 |
+
'bigcode/the-stack-smol-xs',
|
29 |
+
'm-a-p/CodeFeedback-Filtered-Instruction',
|
30 |
+
'jtatman/python-code-dataset-500k',
|
31 |
+
'iamtarun/python_code_instructions_18k_alpaca',
|
32 |
+
'HuggingFaceH4/CodeAlpaca_20K',
|
33 |
+
'cognitivecomputations/dolphin-coder',
|
34 |
+
'fblgit/simple-math',
|
35 |
+
'gair-prox/open-web-math-pro',
|
36 |
+
'rvv-karma/Math-QA',
|
37 |
+
'ajibawa-2023/Maths-College',
|
38 |
+
'microsoft/orca-math-word-problems-200k',
|
39 |
+
'meta-math/MetaMathQA',
|
40 |
+
'TIGER-Lab/MathInstruct',
|
41 |
+
'TIGER-Lab/WebInstructSub',
|
42 |
+
'SkunkworksAI/reasoning-0.01',
|
43 |
+
'KingNish/reasoning-base-20k',
|
44 |
+
'Magpie-Align/Magpie-Reasoning-150K',
|
45 |
+
'thesven/gsm8k-reasoning',
|
46 |
+
'AlgorithmicResearchGroup/math_reasoning_autoformalization_track',
|
47 |
+
'badrex/llm-emoji-dataset',
|
48 |
+
]
|
49 |
+
tags:
|
50 |
+
- litgpt
|
51 |
+
- litdata
|
52 |
---
|
53 |
+
|
54 |
+
# tangled-llama-t-32k-base-v0.1
|
55 |
+
|
56 |
+
![logo](./misc/logo.png)
|
57 |
+
|
58 |
+
A pretrained language model based on the Llama model with about **25M** parameters. This model has been trained on **22.1B** (`22,111,299,936`) tokens from more than **3.6M** (`3,597,088`) dataset rows.
|
59 |
+
|
60 |
+
This model **isn't** designed for immediate use but rather for Continued Pretraining and Finetuning on a downstream task. While it can handle a context length of up to **128K** (`131,072`) tokens, it was pretrained with sequences of **2K** (`2048`) tokens.
|
61 |
+
|
62 |
+
The objective is to streamline the cognitive or reasoning core, eliminating any redundant knowledge from the model.
|
63 |
+
|
64 |
+
[loss, val_loss](https://api.wandb.ai/links/mtasic85/t66yvgeh)
|
65 |
+
|
66 |
+
[val_ppl](https://api.wandb.ai/links/mtasic85/osr62qqd)
|
67 |
+
|
68 |
+
[epoch](https://api.wandb.ai/links/mtasic85/pw0ilz5s)
|
69 |
+
|
70 |
+
[learning_rate](https://api.wandb.ai/links/mtasic85/867ueoyx)
|
71 |
+
|
72 |
+
## lm-evaluation-harness
|
73 |
+
|
74 |
+
```bash
|
75 |
+
litgpt evaluate --tasks 'hellaswag,gsm8k,truthfulqa_mc2,mmlu,winogrande,arc_challenge' --out_dir 'evaluate-quick/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
76 |
+
```
|
77 |
+
|
78 |
+
```bash
|
79 |
+
litgpt evaluate --tasks 'leaderboard' --out_dir 'evaluate-leaderboard/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
80 |
+
```
|
81 |
+
|
82 |
+
```bash
|
83 |
+
litgpt evaluate --tasks 'bbh_zeroshot,bbh_fewshot,bbh_cot_fewshot,bbh_cot_zeroshot' --out_dir 'evaluate-bigbenchhard/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
84 |
+
```
|
85 |
+
|
86 |
+
```bash
|
87 |
+
litgpt evaluate --tasks 'mmlu,mmlu_pro' --out_dir 'evaluate-mmlu/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
88 |
+
```
|
89 |
+
|
90 |
+
```bash
|
91 |
+
litgpt evaluate --tasks 'arc_challenge,boolq,gpqa,hellaswag,openbookqa,piqa,truthfulqa_mc2,winogrande' --out_dir 'evaluate-reasoning/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
92 |
+
```
|
93 |
+
|
94 |
+
```bash
|
95 |
+
litgpt evaluate --tasks 'mmlu_multilingual,mgsm' --out_dir 'evaluate-multilinguals/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
96 |
+
```
|
97 |
+
|
98 |
+
```bash
|
99 |
+
litgpt evaluate --tasks 'gsm8k,mathqa' --out_dir 'evaluate-math/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
100 |
+
```
|
101 |
+
|
102 |
+
```bash
|
103 |
+
litgpt evaluate --tasks 'wikitext,qasper' --out_dir 'evaluate-long/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
|
104 |
+
```
|
config.json
CHANGED
@@ -1,16 +1,16 @@
|
|
1 |
{
|
2 |
-
"_name_or_path": "tangledgroup/tangled-llama-
|
3 |
"architectures": [
|
4 |
"LlamaForCausalLM"
|
5 |
],
|
6 |
"bos_token_id": 1,
|
7 |
"eos_token_id": 2,
|
8 |
-
"hidden_size":
|
9 |
-
"intermediate_size":
|
10 |
"max_position_embeddings": 38400,
|
11 |
"model_type": "llama",
|
12 |
"num_attention_heads": 32,
|
13 |
-
"num_hidden_layers":
|
14 |
"num_key_value_heads": 8,
|
15 |
"rms_norm_eps": 1e-05,
|
16 |
"rope_scaling": null,
|
|
|
1 |
{
|
2 |
+
"_name_or_path": "tangledgroup/tangled-llama-t-128k-base-v0.1",
|
3 |
"architectures": [
|
4 |
"LlamaForCausalLM"
|
5 |
],
|
6 |
"bos_token_id": 1,
|
7 |
"eos_token_id": 2,
|
8 |
+
"hidden_size": 256,
|
9 |
+
"intermediate_size": 1024,
|
10 |
"max_position_embeddings": 38400,
|
11 |
"model_type": "llama",
|
12 |
"num_attention_heads": 32,
|
13 |
+
"num_hidden_layers": 16,
|
14 |
"num_key_value_heads": 8,
|
15 |
"rms_norm_eps": 1e-05,
|
16 |
"rope_scaling": null,
|