File size: 4,212 Bytes
2056078 06d5b88 2056078 06d5b88 2056078 |
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 |
---
base_model: google/gemma-2-2b-it
library_name: peft
license: other
tags:
- llama-factory
- lora
- generated_from_trainer
model-index:
- name: PPRM-gemma-2-2b-it
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# PRM_DPO_GEMMA_ZD_8_18_1
This model is a fine-tuned version of [google/gemma-2-2b-it](https://huggingface.co/google/gemma-2-2b-it) on the prm_dpo dataset.
# Citation
```
@article{zhang2024llama,
title={LLaMA-Berry: Pairwise Optimization for O1-like Olympiad-Level Mathematical Reasoning},
author={Zhang, Di and Wu, Jianbo and Lei, Jingdi and Che, Tong and Li, Jiatong and Xie, Tong and Huang, Xiaoshui and Zhang, Shufei and Pavone, Marco and Li, Yuqiang and others},
journal={arXiv preprint arXiv:2410.02884},
year={2024}
}
@article{zhang2024accessing,
title={Accessing GPT-4 level Mathematical Olympiad Solutions via Monte Carlo Tree Self-refine with LLaMa-3 8B},
author={Zhang, Di and Li, Jiatong and Huang, Xiaoshui and Zhou, Dongzhan and Li, Yuqiang and Ouyang, Wanli},
journal={arXiv preprint arXiv:2406.07394},
year={2024}
}
```
## Model usage
`server.py`
```
import json
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
# Initialize FastAPI
app = FastAPI()
# Device configuration
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Model and tokenizer loading (as you provided)
model_name = "google/gemma-2-2b-it"
lora_checkpoint_path = "qq8933/PPRM-gemma-2-2b-it"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map='cuda')
model = PeftModel.from_pretrained(base_model, lora_checkpoint_path, device_map='cuda')
yes_token_id = tokenizer.convert_tokens_to_ids("yes")
no_token_id = tokenizer.convert_tokens_to_ids("no")
# Request model
class InputRequest(BaseModel):
text: str
# Predict function
def predict(qeustion,answer_1,answer_2):
prompt_template = """Problem:\n\n{}\n\nFirst Answer:\n\n{}\n\nSecond Answer:\n\n{}\n\nIs First Answer better than Second Answer?\n\n"""
input_text = prompt_template.format(qeustion,answer_1,answer_2)
input_text = tokenizer.apply_chat_template(
[{'role': 'user', 'content': input_text}], tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(input_text, return_tensors="pt").to(device)
with torch.no_grad():
generated_outputs = model.generate(
**inputs, max_new_tokens=2, output_scores=True, return_dict_in_generate=True
)
scores = generated_outputs.scores
first_token_logits = scores[0]
yes_logit = first_token_logits[0, yes_token_id].item()
no_logit = first_token_logits[0, no_token_id].item()
return {
"yes_logit": yes_logit,
"no_logit": no_logit,
"logit_difference": yes_logit - no_logit
}
# Define API endpoint
@app.post("/predict")
async def get_prediction(input_request: InputRequest):
payload = json.loads(input_request.text)
qeustion,answer_1,answer_2 = payload['qeustion'],payload['answer_1'],payload['answer_2']
try:
result = predict(qeustion,answer_1,answer_2)
return result
except Exception as e:
raise HTTPException(status_code=500, detail=str(e))
```
```
uvicorn server:app --host 0.0.0.0 --port $MASTER_PORT --workers 1
```
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 4
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 16
- gradient_accumulation_steps: 2
- total_train_batch_size: 128
- total_eval_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 1.0
### Framework versions
- PEFT 0.11.1
- Transformers 4.44.0
- Pytorch 2.3.1
- Datasets 2.20.0
- Tokenizers 0.19.1 |