File size: 4,420 Bytes
4d9810e
833888d
 
 
ea9a4cd
2a74db6
 
833888d
 
 
ea9a4cd
 
 
4d9810e
833888d
 
 
ea9a4cd
833888d
2a74db6
 
833888d
f0db376
 
 
833888d
 
ea9a4cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
833888d
cbe7f49
 
 
833888d
 
ea9a4cd
833888d
 
 
9877e68
4d9810e
833888d
4d9810e
ea9a4cd
 
 
 
 
 
 
 
 
 
 
833888d
 
 
 
 
 
 
 
 
 
9877e68
4d9810e
2a74db6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
833888d
4d9810e
ea9a4cd
833888d
 
 
ea9a4cd
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
---
base_model: meta-llama/Llama-2-7b-chat-hf
tags:
- generated_from_trainer
- trl
metrics:
- accuracy
model-index:
- name: llama-2-7b-reward-oasst1
  results: []
datasets:
- tasksource/oasst1_pairwise_rlhf_reward
library_name: peft
---

# llama-2-7b-reward-oasst1

This model is a fine-tuned version of [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) on the first 10000 rows of the [tasksource/oasst1_pairwise_rlhf_reward](https://huggingface.co/datasets/tasksource/oasst1_pairwise_rlhf_reward) dataset.
It achieves the following results on the evaluation set:
- Loss: 0.5713
- Accuracy: 0.7435

See also [vincentmin/llama-2-13b-reward-oasst1](https://huggingface.co/vincentmin/llama-2-13b-reward-oasst1) for a 13b version of this model.


## Model description

This is a reward model trained with QLoRA in 4bit precision. The base model is [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) for which you need to have accepted the license in order to be able use it. Once you've been given permission, you can load the reward model as follows:
```
import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSequenceClassification, AutoTokenizer

peft_model_id = "vincentmin/llama-2-7b-reward-oasst1"
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForSequenceClassification.from_pretrained(
    config.base_model_name_or_path,
    num_labels=1,
    load_in_4bit=True,
    torch_dtype=torch.float16,
)
model = PeftModel.from_pretrained(model, peft_model_id)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path, use_auth_token=True)
model.eval()
with torch.no_grad():
  reward = model(**tokenizer("prompter: hello world. assistant: foo bar", return_tensors='pt')).logits
reward
```
For best results, one should use the prompt format used during training:
```
prompt = "prompter: <prompt_1> assistant: <response_1> prompter: <prompt_2> ..."
```

Please use a version of peft where [#755](https://github.com/huggingface/peft/pull/755) has been merged to make sure the model is loaded correctly. You can install `peft` with `pip install git+https://github.com/huggingface/peft.git` to make sure this is the case.


## Intended uses & limitations

Since the model was trained on oasst1 data, the reward will reflect any biases present in the oasst1 data.

## Training and evaluation data

The model was trained using QLoRA and the `trl` library's `RewardTrainer` on the [tasksource/oasst1_pairwise_rlhf_reward](https://huggingface.co/datasets/tasksource/oasst1_pairwise_rlhf_reward) dataset. Examples with more than 1024 tokens were filtered out and the training data was restricted to the first 10000 rows of the filtered dataset.

### Training hyperparameters

The following `bitsandbytes` quantization config was used during training:
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: False
- bnb_4bit_compute_dtype: float16

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 1
- max_seq_length: 1024

### Training results

| Training Loss | Epoch | Step | Validation Loss | Accuracy |
|:-------------:|:-----:|:----:|:---------------:|:--------:|
| 0.8409        | 0.1   | 250  | 0.8243          | 0.6220   |
| 0.6288        | 0.2   | 500  | 0.7539          | 0.6715   |
| 0.5882        | 0.3   | 750  | 0.6792          | 0.7075   |
| 0.7671        | 0.4   | 1000 | 0.6130          | 0.7334   |
| 0.5782        | 0.5   | 1250 | 0.6115          | 0.7255   |
| 0.5691        | 0.6   | 1500 | 0.5795          | 0.7413   |
| 0.6579        | 0.7   | 1750 | 0.5774          | 0.7469   |
| 0.6107        | 0.8   | 2000 | 0.5691          | 0.7402   |
| 0.6255        | 0.9   | 2250 | 0.5710          | 0.7435   |
| 0.7034        | 1.0   | 2500 | 0.5713          | 0.7435   |


### Framework versions

- PEFT 0.5.0.dev0 (with https://github.com/huggingface/peft/pull/755)
- Transformers 4.32.0.dev0
- Pytorch 2.0.1+cu118
- Datasets 2.14.0
- Tokenizers 0.13.3