Hugging Face's logo

qq8933
/
PPRM-gemma-2-2b-it

PEFT
Safetensors
llama-factory
lora
Generated from Trainer
Model card Files Community
PPRM-gemma-2-2b-it
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