Commit
•
ce3900f
1
Parent(s):
095a04f
uploaded readme
Browse files
README.md
ADDED
|
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Quantization made by Richard Erkhov.
|
| 2 |
+
|
| 3 |
+
[Github](https://github.com/RichardErkhov)
|
| 4 |
+
|
| 5 |
+
[Discord](https://discord.gg/pvy7H8DZMG)
|
| 6 |
+
|
| 7 |
+
[Request more models](https://github.com/RichardErkhov/quant_request)
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
pair-preference-model-LLaMA3-8B - GGUF
|
| 11 |
+
- Model creator: https://huggingface.co/RLHFlow/
|
| 12 |
+
- Original model: https://huggingface.co/RLHFlow/pair-preference-model-LLaMA3-8B/
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
| Name | Quant method | Size |
|
| 16 |
+
| ---- | ---- | ---- |
|
| 17 |
+
| [pair-preference-model-LLaMA3-8B.Q2_K.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q2_K.gguf) | Q2_K | 2.96GB |
|
| 18 |
+
| [pair-preference-model-LLaMA3-8B.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.IQ3_XS.gguf) | IQ3_XS | 3.28GB |
|
| 19 |
+
| [pair-preference-model-LLaMA3-8B.IQ3_S.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.IQ3_S.gguf) | IQ3_S | 3.43GB |
|
| 20 |
+
| [pair-preference-model-LLaMA3-8B.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q3_K_S.gguf) | Q3_K_S | 3.41GB |
|
| 21 |
+
| [pair-preference-model-LLaMA3-8B.IQ3_M.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.IQ3_M.gguf) | IQ3_M | 3.52GB |
|
| 22 |
+
| [pair-preference-model-LLaMA3-8B.Q3_K.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q3_K.gguf) | Q3_K | 3.74GB |
|
| 23 |
+
| [pair-preference-model-LLaMA3-8B.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q3_K_M.gguf) | Q3_K_M | 3.74GB |
|
| 24 |
+
| [pair-preference-model-LLaMA3-8B.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q3_K_L.gguf) | Q3_K_L | 4.03GB |
|
| 25 |
+
| [pair-preference-model-LLaMA3-8B.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.IQ4_XS.gguf) | IQ4_XS | 4.18GB |
|
| 26 |
+
| [pair-preference-model-LLaMA3-8B.Q4_0.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q4_0.gguf) | Q4_0 | 4.34GB |
|
| 27 |
+
| [pair-preference-model-LLaMA3-8B.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.IQ4_NL.gguf) | IQ4_NL | 4.38GB |
|
| 28 |
+
| [pair-preference-model-LLaMA3-8B.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q4_K_S.gguf) | Q4_K_S | 4.37GB |
|
| 29 |
+
| [pair-preference-model-LLaMA3-8B.Q4_K.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q4_K.gguf) | Q4_K | 4.58GB |
|
| 30 |
+
| [pair-preference-model-LLaMA3-8B.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q4_K_M.gguf) | Q4_K_M | 4.58GB |
|
| 31 |
+
| [pair-preference-model-LLaMA3-8B.Q4_1.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q4_1.gguf) | Q4_1 | 4.78GB |
|
| 32 |
+
| [pair-preference-model-LLaMA3-8B.Q5_0.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q5_0.gguf) | Q5_0 | 5.21GB |
|
| 33 |
+
| [pair-preference-model-LLaMA3-8B.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q5_K_S.gguf) | Q5_K_S | 5.21GB |
|
| 34 |
+
| [pair-preference-model-LLaMA3-8B.Q5_K.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q5_K.gguf) | Q5_K | 5.34GB |
|
| 35 |
+
| [pair-preference-model-LLaMA3-8B.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q5_K_M.gguf) | Q5_K_M | 5.34GB |
|
| 36 |
+
| [pair-preference-model-LLaMA3-8B.Q5_1.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q5_1.gguf) | Q5_1 | 5.65GB |
|
| 37 |
+
| [pair-preference-model-LLaMA3-8B.Q6_K.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q6_K.gguf) | Q6_K | 6.14GB |
|
| 38 |
+
| [pair-preference-model-LLaMA3-8B.Q8_0.gguf](https://huggingface.co/RichardErkhov/RLHFlow_-_pair-preference-model-LLaMA3-8B-gguf/blob/main/pair-preference-model-LLaMA3-8B.Q8_0.gguf) | Q8_0 | 7.95GB |
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
Original model description:
|
| 44 |
+
---
|
| 45 |
+
license: llama3
|
| 46 |
+
---
|
| 47 |
+
This preference model is trained from [LLaMA3-8B-it](meta-llama/Meta-Llama-3-8B-Instruct) with the training script at [Reward Modeling](https://github.com/RLHFlow/RLHF-Reward-Modeling/tree/pm_dev/pair-pm).
|
| 48 |
+
|
| 49 |
+
The dataset is RLHFlow/pair_preference_model_dataset. It achieves Chat-98.6, Char-hard 65.8, Safety 89.6, and reasoning 94.9 in reward bench.
|
| 50 |
+
|
| 51 |
+
See our paper [RLHF Workflow: From Reward Modeling to Online RLHF](https://arxiv.org/abs/2405.07863) for more details of this model.
|
| 52 |
+
|
| 53 |
+
## Service the RM
|
| 54 |
+
|
| 55 |
+
Here is an example to use the Preference Model to rank a pair. For n>2 responses, it is recommened to use the tournament style ranking strategy to get the best response so that the complexity is linear in n.
|
| 56 |
+
|
| 57 |
+
```python
|
| 58 |
+
device = 0
|
| 59 |
+
|
| 60 |
+
model = AutoModelForCausalLM.from_pretrained(script_args.preference_name_or_path,
|
| 61 |
+
torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2").cuda()
|
| 62 |
+
tokenizer = AutoTokenizer.from_pretrained(script_args.preference_name_or_path, use_fast=True)
|
| 63 |
+
tokenizer_plain = AutoTokenizer.from_pretrained(script_args.preference_name_or_path, use_fast=True)
|
| 64 |
+
tokenizer_plain.chat_template = "\n{% for message in messages %}{% if loop.index0 % 2 == 0 %}\n\n<turn> user\n {{ message['content'] }}{% else %}\n\n<turn> assistant\n {{ message['content'] }}{% endif %}{% endfor %}\n\n\n"
|
| 65 |
+
|
| 66 |
+
prompt_template = "[CONTEXT] {context} [RESPONSE A] {response_A} [RESPONSE B] {response_B} \n"
|
| 67 |
+
token_id_A = tokenizer.encode("A", add_special_tokens=False)
|
| 68 |
+
token_id_B = tokenizer.encode("B", add_special_tokens=False)
|
| 69 |
+
assert len(token_id_A) == 1 and len(token_id_B) == 1
|
| 70 |
+
token_id_A = token_id_A[0]
|
| 71 |
+
token_id_B = token_id_B[0]
|
| 72 |
+
temperature = 1.0
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
model.eval()
|
| 76 |
+
response_chosen = "BBBB"
|
| 77 |
+
response_rejected = "CCCC"
|
| 78 |
+
|
| 79 |
+
## We can also handle multi-turn conversation.
|
| 80 |
+
instruction = [{"role": "user", "content": ...},
|
| 81 |
+
{"role": "assistant", "content": ...},
|
| 82 |
+
{"role": "user", "content": ...},
|
| 83 |
+
]
|
| 84 |
+
context = tokenizer_plain.apply_chat_template(instruction, tokenize=False)
|
| 85 |
+
responses = [response_chosen, response_rejected]
|
| 86 |
+
probs_chosen = []
|
| 87 |
+
|
| 88 |
+
for chosen_position in [0, 1]:
|
| 89 |
+
# we swap order to mitigate position bias
|
| 90 |
+
response_A = responses[chosen_position]
|
| 91 |
+
response_B = responses[1 - chosen_position]
|
| 92 |
+
prompt = prompt_template.format(context=context, response_A=response_A, response_B=response_B)
|
| 93 |
+
message = [
|
| 94 |
+
{"role": "user", "content": prompt},
|
| 95 |
+
]
|
| 96 |
+
|
| 97 |
+
input_ids = tokenizer.encode(tokenizer.apply_chat_template(message, tokenize=False).replace(tokenizer.bos_token, ""), return_tensors='pt', add_special_tokens=False).cuda()
|
| 98 |
+
|
| 99 |
+
with torch.no_grad():
|
| 100 |
+
output = model(input_ids)
|
| 101 |
+
logit_A = output.logits[0, -1, token_id_A].item()
|
| 102 |
+
logit_B = output.logits[0, -1, token_id_B].item()
|
| 103 |
+
# take softmax to get the probability; using numpy
|
| 104 |
+
Z = np.exp(logit_A / temperature) + np.exp(logit_B / temperature)
|
| 105 |
+
logit_chosen = [logit_A, logit_B][chosen_position]
|
| 106 |
+
prob_chosen = np.exp(logit_chosen / temperature) / Z
|
| 107 |
+
probs_chosen.append(prob_chosen)
|
| 108 |
+
|
| 109 |
+
avg_prob_chosen = np.mean(probs_chosen)
|
| 110 |
+
correct = 0.5 if avg_prob_chosen == 0.5 else float(avg_prob_chosen > 0.5)
|
| 111 |
+
print(correct)
|
| 112 |
+
```
|
| 113 |
+
|
| 114 |
+
## Citation
|
| 115 |
+
If you use this model in your research, please consider citing our paper
|
| 116 |
+
```
|
| 117 |
+
@misc{rlhflow,
|
| 118 |
+
title={RLHF Workflow: From Reward Modeling to Online RLHF},
|
| 119 |
+
author={Hanze Dong and Wei Xiong and Bo Pang and Haoxiang Wang and Han Zhao and Yingbo Zhou and Nan Jiang and Doyen Sahoo and Caiming Xiong and Tong Zhang},
|
| 120 |
+
year={2024},
|
| 121 |
+
eprint={2405.07863},
|
| 122 |
+
archivePrefix={arXiv},
|
| 123 |
+
primaryClass={cs.LG}
|
| 124 |
+
}
|
| 125 |
+
```
|
| 126 |
+
and Google's Slic paper (which initially proposes this pairwise preference model)
|
| 127 |
+
```
|
| 128 |
+
@article{zhao2023slic,
|
| 129 |
+
title={Slic-hf: Sequence likelihood calibration with human feedback},
|
| 130 |
+
author={Zhao, Yao and Joshi, Rishabh and Liu, Tianqi and Khalman, Misha and Saleh, Mohammad and Liu, Peter J},
|
| 131 |
+
journal={arXiv preprint arXiv:2305.10425},
|
| 132 |
+
year={2023}
|
| 133 |
+
}
|
| 134 |
+
```
|
| 135 |
+
|