Hugging Face's logo

RichardErkhov
/
UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf

GGUF
Inference Endpoints
conversational
Model card Files Community
RichardErkhov commited on
Commit
3a4a9a6
·
verified ·
1 Parent(s): 20f8fc3

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +138 -0
README.md ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ Mistral7B-PairRM-SPPO-Iter2 - GGUF
11
+ - Model creator: https://huggingface.co/UCLA-AGI/
12
+ - Original model: https://huggingface.co/UCLA-AGI/Mistral7B-PairRM-SPPO-Iter2/
13
+
14
+
15
+ | Name | Quant method | Size |
16
+ | ---- | ---- | ---- |
17
+ | [Mistral7B-PairRM-SPPO-Iter2.Q2_K.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q2_K.gguf) | Q2_K | 2.53GB |
18
+ | [Mistral7B-PairRM-SPPO-Iter2.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.IQ3_XS.gguf) | IQ3_XS | 2.81GB |
19
+ | [Mistral7B-PairRM-SPPO-Iter2.IQ3_S.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.IQ3_S.gguf) | IQ3_S | 2.96GB |
20
+ | [Mistral7B-PairRM-SPPO-Iter2.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q3_K_S.gguf) | Q3_K_S | 2.95GB |
21
+ | [Mistral7B-PairRM-SPPO-Iter2.IQ3_M.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.IQ3_M.gguf) | IQ3_M | 3.06GB |
22
+ | [Mistral7B-PairRM-SPPO-Iter2.Q3_K.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q3_K.gguf) | Q3_K | 3.28GB |
23
+ | [Mistral7B-PairRM-SPPO-Iter2.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q3_K_M.gguf) | Q3_K_M | 3.28GB |
24
+ | [Mistral7B-PairRM-SPPO-Iter2.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q3_K_L.gguf) | Q3_K_L | 3.56GB |
25
+ | [Mistral7B-PairRM-SPPO-Iter2.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.IQ4_XS.gguf) | IQ4_XS | 3.67GB |
26
+ | [Mistral7B-PairRM-SPPO-Iter2.Q4_0.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q4_0.gguf) | Q4_0 | 3.83GB |
27
+ | [Mistral7B-PairRM-SPPO-Iter2.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.IQ4_NL.gguf) | IQ4_NL | 3.87GB |
28
+ | [Mistral7B-PairRM-SPPO-Iter2.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q4_K_S.gguf) | Q4_K_S | 3.86GB |
29
+ | [Mistral7B-PairRM-SPPO-Iter2.Q4_K.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q4_K.gguf) | Q4_K | 4.07GB |
30
+ | [Mistral7B-PairRM-SPPO-Iter2.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q4_K_M.gguf) | Q4_K_M | 4.07GB |
31
+ | [Mistral7B-PairRM-SPPO-Iter2.Q4_1.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q4_1.gguf) | Q4_1 | 4.24GB |
32
+ | [Mistral7B-PairRM-SPPO-Iter2.Q5_0.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q5_0.gguf) | Q5_0 | 4.65GB |
33
+ | [Mistral7B-PairRM-SPPO-Iter2.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q5_K_S.gguf) | Q5_K_S | 4.65GB |
34
+ | [Mistral7B-PairRM-SPPO-Iter2.Q5_K.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q5_K.gguf) | Q5_K | 4.78GB |
35
+ | [Mistral7B-PairRM-SPPO-Iter2.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q5_K_M.gguf) | Q5_K_M | 4.78GB |
36
+ | [Mistral7B-PairRM-SPPO-Iter2.Q5_1.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q5_1.gguf) | Q5_1 | 5.07GB |
37
+ | [Mistral7B-PairRM-SPPO-Iter2.Q6_K.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q6_K.gguf) | Q6_K | 5.53GB |
38
+ | [Mistral7B-PairRM-SPPO-Iter2.Q8_0.gguf](https://huggingface.co/RichardErkhov/UCLA-AGI_-_Mistral7B-PairRM-SPPO-Iter2-gguf/blob/main/Mistral7B-PairRM-SPPO-Iter2.Q8_0.gguf) | Q8_0 | 7.17GB |
39
+
40
+
41
+
42
+
43
+ Original model description:
44
+ ---
45
+ license: apache-2.0
46
+ datasets:
47
+ - openbmb/UltraFeedback
48
+ language:
49
+ - en
50
+ pipeline_tag: text-generation
51
+ ---
52
+ Self-Play Preference Optimization for Language Model Alignment (https://arxiv.org/abs/2405.00675)
53
+
54
+ # Mistral7B-PairRM-SPPO-Iter2
55
+
56
+ This model was developed using [Self-Play Preference Optimization](https://arxiv.org/abs/2405.00675) at iteration 2, based on the [mistralai/Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2) architecture as starting point. We utilized the prompt sets from the [openbmb/UltraFeedback](https://huggingface.co/datasets/openbmb/UltraFeedback) dataset, splited to 3 parts for 3 iterations by [snorkelai/Snorkel-Mistral-PairRM-DPO-Dataset](https://huggingface.co/datasets/snorkelai/Snorkel-Mistral-PairRM-DPO-Dataset). All responses used are synthetic.
57
+
58
+ **This is the model reported in the paper** , with K=5 (generate 5 responses per iteration). We attached the Arena-Hard eval results in this model page.
59
+
60
+ ## Links to Other Models
61
+ - [Mistral7B-PairRM-SPPO-Iter1](https://huggingface.co/UCLA-AGI/Mistral7B-PairRM-SPPO-Iter1)
62
+ - [Mistral7B-PairRM-SPPO-Iter2](https://huggingface.co/UCLA-AGI/Mistral7B-PairRM-SPPO-Iter2)
63
+ - [Mistral7B-PairRM-SPPO-Iter3](https://huggingface.co/UCLA-AGI/Mistral7B-PairRM-SPPO-Iter3)
64
+ - [Mistral7B-PairRM-SPPO](https://huggingface.co/UCLA-AGI/Mistral7B-PairRM-SPPO)
65
+
66
+
67
+ ### Model Description
68
+
69
+ - Model type: A 7B parameter GPT-like model fine-tuned on synthetic datasets.
70
+ - Language(s) (NLP): Primarily English
71
+ - License: Apache-2.0
72
+ - Finetuned from model: mistralai/Mistral-7B-Instruct-v0.2
73
+
74
+
75
+ ## [AlpacaEval Leaderboard Evaluation Results](https://tatsu-lab.github.io/alpaca_eval/)
76
+
77
+
78
+ | Model | LC. Win Rate | Win Rate | Avg. Length |
79
+ |-------------------------------------------|:------------:|:--------:|:-----------:|
80
+ | Mistral7B-PairRM-SPPO Iter 1 | 24.79 | 23.51 | 1855 |
81
+ | Mistral7B-PairRM-SPPO Iter 2 | 26.89 | 27.62 | 2019 |
82
+ | Mistral7B-PairRM-SPPO Iter 3 | 28.53 | 31.02 | 2163 |
83
+ | Mistral7B-PairRM-SPPO Iter 1 (best-of-16) | 28.71 | 27.77 | 1901 |
84
+ | Mistral7B-PairRM-SPPO Iter 2 (best-of-16) | 31.23 | 32.12 | 2035 |
85
+ | Mistral7B-PairRM-SPPO Iter 3 (best-of-16) | 32.13 | 34.94 | 2174 |
86
+ ## [Arena-Hard Evaluation Results](https://github.com/lm-sys/arena-hard)
87
+
88
+ Model | Score | 95% CI | average \# Tokens |
89
+ |----------|-----------|--------------|-----------|
90
+ Mistral7B-PairRM-SPPO-Iter3| 23.3 | (-1.8, 1.8)|578|
91
+
92
+ ## [Open LLM Leaderboard Evaluation Results](https://github.com/EleutherAI/lm-evaluation-harness)
93
+
94
+ Results are reported by using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) v0.4.1
95
+
96
+ | | arc_challenge | truthfulqa_mc2 | winogrande | gsm8k | hellaswag | mmlu | average |
97
+ |--------|---------------|----------------|------------|-------|-----------|-------|---------|
98
+ | Mistral7B-PairRM-SPPO Iter 1 | 65.02 | 69.4 | 77.82 | 43.82 | 85.11 | 58.84 | 66.67 |
99
+ | Mistral7B-PairRM-SPPO Iter 2 | 65.53 | 69.55 | 77.03 | 44.35 | 85.29 | 58.72 | 66.75 |
100
+ | Mistral7B-PairRM-SPPO Iter 3 | 65.36 | 69.97 | 76.8 | 42.68 | 85.16 | 58.45 | 66.4 |
101
+ ## [MT-Bench Evaluation Results](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge)
102
+
103
+ | | 1st Turn | 2nd Turn | Average |
104
+ |--------|----------|----------|---------|
105
+ | Mistral7B-PairRM-SPPO Iter 1 | 7.63 | 6.79 | 7.21 |
106
+ | Mistral7B-PairRM-SPPO Iter 2 | 7.90 | 7.08 | 7.49 |
107
+ | Mistral7B-PairRM-SPPO Iter 3 | 7.84 | 7.34 | 7.59 |
108
+
109
+ ### Training hyperparameters
110
+ The following hyperparameters were used during training:
111
+
112
+ - learning_rate: 5e-07
113
+ - eta: 1000
114
+ - per_device_train_batch_size: 8
115
+ - gradient_accumulation_steps: 1
116
+ - seed: 42
117
+ - distributed_type: deepspeed_zero3
118
+ - num_devices: 8
119
+ - optimizer: RMSProp
120
+ - lr_scheduler_type: linear
121
+ - lr_scheduler_warmup_ratio: 0.1
122
+ - num_train_epochs: 18.0 (stop at epoch=1.0)
123
+
124
+
125
+
126
+
127
+ ## Citation
128
+ ```
129
+ @misc{wu2024self,
130
+ title={Self-Play Preference Optimization for Language Model Alignment},
131
+ author={Wu, Yue and Sun, Zhiqing and Yuan, Huizhuo and Ji, Kaixuan and Yang, Yiming and Gu, Quanquan},
132
+ year={2024},
133
+ eprint={2405.00675},
134
+ archivePrefix={arXiv},
135
+ primaryClass={cs.LG}
136
+ }
137
+ ```
138
+