File size: 10,109 Bytes
16df845 2cece3e 16df845 e819930 16df845 |
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 |
---
language:
- en
license: other
license_name: qwen
license_link: https://huggingface.co./Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE
library_name: transformers
tags:
- generated_from_trainer
base_model: fblgit/miniclaus-qw1.5B-UNAMGS
model-index:
- name: miniclaus-qw1.5B-UNAMGS
results: []
datasets:
- openai/gsm8k
---
# miniclaus-qw1.5B-UNAMGS-GRPO
This version is RL with GRPO on GSM8k for 1400 steps using this code:
```
# train_grpo.py
import re
import torch
from datasets import load_dataset, Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import LoraConfig
from trl import GRPOConfig, GRPOTrainer
# Load and prep dataset
SYSTEM_PROMPT = """
Respond exclusively the following format:
<reasoning>
...
</reasoning>
<answer>
...
</answer>
Its imperative to follow strictritly your final result answer within the <answer>$result</answer> and be terse.
"""
XML_COT_FORMAT = """\
<reasoning>
{reasoning}
</reasoning>
<answer>
{answer}
</answer>
"""
def extract_xml_answer(text: str) -> str:
answer = text.split("<answer>")[-1]
answer = answer.split("</answer>")[0]
return answer.strip()
def extract_hash_answer(text: str) -> str | None:
if "####" not in text:
return None
return text.split("####")[1].strip()
# uncomment middle messages for 1-shot prompting
def get_gsm8k_questions(split = "train") -> Dataset:
data = load_dataset('openai/gsm8k', 'main')[split] # type: ignore
data = data.map(lambda x: { # type: ignore
'prompt': [
{'role': 'system', 'content': SYSTEM_PROMPT},
#{'role': 'user', 'content': 'What is the largest single-digit prime number?'},
#{'role': 'assistant', 'content': XML_COT_FORMAT.format(
# reasoning="9 is divisble by 3 and 8 is divisible by 2, but 7 is prime.",
# answer="7"
#)},
{'role': 'user', 'content': x['question']}
],
'answer': extract_hash_answer(x['answer'])
}) # type: ignore
return data # type: ignore
dataset = get_gsm8k_questions()
# Reward functions
def int_reward_func(completions, **kwargs) -> list[float]:
responses = [completion[0]['content'] for completion in completions]
extracted_responses = [extract_xml_answer(r) for r in responses]
return [0.5 if r.isdigit() else 0.0 for r in extracted_responses]
def strict_format_reward_func(completions, **kwargs) -> list[float]:
"""Reward function that checks if the completion has a specific format."""
pattern = r"^<reasoning>\n.*?\n</reasoning>\n<answer>\n.*?\n</answer>\n$"
responses = [completion[0]["content"] for completion in completions]
matches = [re.match(pattern, r) for r in responses]
return [0.5 if match else 0.0 for match in matches]
def soft_format_reward_func(completions, **kwargs) -> list[float]:
"""Reward function that checks if the completion has a specific format."""
pattern = r"<reasoning>.*?</reasoning>\s*<answer>.*?</answer>"
responses = [completion[0]["content"] for completion in completions]
matches = [re.match(pattern, r) for r in responses]
return [0.5 if match else 0.0 for match in matches]
def correctness_reward_func(prompts, completions, answer, **kwargs) -> list[float]:
responses = [completion[0]['content'] for completion in completions]
q = prompts[0][-1]['content']
extracted_responses = [extract_xml_answer(r) for r in responses]
# Extract the last number from each extracted response
last_numbers = []
for response in extracted_responses:
numbers = re.findall(r'\d+', response)
last_num = numbers[-1] if numbers else None
last_numbers.append(last_num)
print('-'*20, f"Question:\n{q}", f"\nAnswer:\n{answer[0]}", f"\nResponse:\n{responses[0]}",
f"\nExtracted:\n{extracted_responses[0]}", f"\nLast Number:\n{last_numbers[0]}")
# Compare the last number to the answer
return [2.0 if ln == a else 0.0 for ln, a in zip(last_numbers, answer)]
def count_xml(text) -> float:
count = 0.0
if text.count("<reasoning>\n") == 1:
count += 0.125
if text.count("\n</reasoning>\n") == 1:
count += 0.125
if text.count("\n<answer>\n") == 1:
count += 0.125
count -= len(text.split("\n</answer>\n")[-1])*0.001
if text.count("\n</answer>") == 1:
count += 0.125
count -= (len(text.split("\n</answer>")[-1]) - 1)*0.001
return count
def xmlcount_reward_func(completions, **kwargs) -> list[float]:
contents = [completion[0]["content"] for completion in completions]
return [count_xml(c) for c in contents]
model_name = 'fblgit/miniclaus-qw1.5B-UNAMGS'
if "Llama" in model_name or 'l318b' in model_name:
output_dir = "outputs/Llama-1B-GRPO"
run_name = "Llama-1B-GRPO-gsm8k"
else:
output_dir="outputs/Qwen-1.5B-GRPO"
run_name="Qwen-1.5B-GRPO-gsm8k"
training_args = GRPOConfig(
output_dir=output_dir,
run_name=run_name,
learning_rate=5e-6,
adam_beta1 = 0.9,
adam_beta2 = 0.99,
weight_decay = 0.1,
warmup_ratio = 0.1,
lr_scheduler_type='cosine',
logging_steps=1,
bf16=True,
tf32=True,
per_device_train_batch_size=1,
gradient_accumulation_steps=4,
#num_generations=16,
num_generations=6,
max_prompt_length=256,
max_completion_length=512,
num_train_epochs=1,
save_steps=100,
max_grad_norm=0.1,
report_to="wandb",
log_on_each_node=False,
)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
device_map='cuda:0',
use_cache=True,
).to(device="cuda:0", dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
# use peft at your own risk; not working for me with multi-GPU training
trainer = GRPOTrainer(
model=model,
processing_class=tokenizer,
reward_funcs=[
xmlcount_reward_func,
soft_format_reward_func,
strict_format_reward_func,
int_reward_func,
correctness_reward_func],
args=training_args,
train_dataset=dataset,
)
trainer.train()
```
Trained with `Magpie-Align/Magpie-Pro-MT-300K-v0.1` and `GSM8k`
Using MGS & UNA (MLP) on this tiny but powerful model, together with GRPO.
[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
## Benchmarks
So far we ran a few:
```
| Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr|
|-------------------------------------|-------|------|-----:|--------|---|-----:|---|-----:|
|leaderboard_gpqa | N/A| | | | | | | |
| - leaderboard_gpqa_diamond | 1|none | 0|acc_norm|↑ |0.3030|± |0.0327|
| - leaderboard_gpqa_extended | 1|none | 0|acc_norm|↑ |0.3004|± |0.0196|
| - leaderboard_gpqa_main | 1|none | 0|acc_norm|↑ |0.2969|± |0.0216|
|leaderboard_musr | N/A| | | | | | | |
| - leaderboard_musr_murder_mysteries | 1|none | 0|acc_norm|↑ |0.5400|± |0.0316|
| - leaderboard_musr_object_placements| 1|none | 0|acc_norm|↑ |0.3203|± |0.0292|
| - leaderboard_musr_team_allocation | 1|none | 0|acc_norm|↑ |0.4080|± |0.0311|
|Tasks|Version| Filter |n-shot| Metric | |Value | |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.5974|± |0.0135|
| | |strict-match | 5|exact_match|↑ |0.5921|± |0.0135|
```
There is some increased score in GSM and GPQA & MUSR, but this doesnt happens in all checkpoints and this is the one with the best marks.
## Thanks
- Deepseek Team for the GRPO researches
- HuggingFace for adopting GRPO on TRL
- Qwen Team for their outstanding model
- MagPie Team for contributing plenty of datasets
- Cybertron Cloud Compute
## Citations
```
@misc{miniclaus-qw15,
title={MiniClaus: 1.5B UNAMGS},
author={Xavier Murias},
year={2024},
publisher = {HuggingFace},
journal = {HuggingFace repository},
howpublished = {\url{https://huggingface.co./fblgit/miniclaus-qw1.5B-UNAMGS}},
}
@misc{Magpie,
title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing},
author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
year={2024},
eprint={2406.08464},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@misc{qwen2.5,
title = {Qwen2.5: A Party of Foundation Models},
url = {https://qwenlm.github.io/blog/qwen2.5/},
author = {Qwen Team},
month = {September},
year = {2024}
}
@article{qwen2,
title={Qwen2 Technical Report},
author={An Yang and Baosong Yang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Zhou and Chengpeng Li and Chengyuan Li and Dayiheng Liu and Fei Huang and Guanting Dong and Haoran Wei and Huan Lin and Jialong Tang and Jialin Wang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Ma and Jin Xu and Jingren Zhou and Jinze Bai and Jinzheng He and Junyang Lin and Kai Dang and Keming Lu and Keqin Chen and Kexin Yang and Mei Li and Mingfeng Xue and Na Ni and Pei Zhang and Peng Wang and Ru Peng and Rui Men and Ruize Gao and Runji Lin and Shijie Wang and Shuai Bai and Sinan Tan and Tianhang Zhu and Tianhao Li and Tianyu Liu and Wenbin Ge and Xiaodong Deng and Xiaohuan Zhou and Xingzhang Ren and Xinyu Zhang and Xipin Wei and Xuancheng Ren and Yang Fan and Yang Yao and Yichang Zhang and Yu Wan and Yunfei Chu and Yuqiong Liu and Zeyu Cui and Zhenru Zhang and Zhihao Fan},
journal={arXiv preprint arXiv:2407.10671},
year={2024}
}
``` |