Open-ended instruction following on Vicuna Eval v1.3 (test)
65A Win RateBPO + Vicuna-v1.3 7B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BPO + Vicuna-v1.3 7BBase LLM=Vicuna-v1.3, Model A=7B + BPO, Model B=7B, Evaluation LLM=GPT-42023.11 | 65 | 8.7 | 26.3 | 18.5 | |
| BPO + Llama-2-chat 13BBase LLM=Llama-2-chat, Model A=13B + BPO, Model B=13B, Evaluation LLM=GPT-42023.11 | 61.3 | 2.5 | 36.2 | 18.1 | |
| BPO + Llama-2-chat 13B (Cross-size)Base LLM=Llama-2-chat, Model A=13B + BPO, Model B=70B, Evaluation LLM=GPT-42023.11 | 61.3 | 0 | 38.7 | 11.9 | |
| BPO + Llama-2-chat 7BBase LLM=Llama-2-chat, Model A=7B + BPO, Model B=7B, Evaluation LLM=GPT-42023.11 | 60 | 2.5 | 37.5 | 17.4 | |
| BPO + Llama-2-chat 70BBase LLM=Llama-2-chat, Model A=70B + BPO, Model B=70B, Evaluation LLM=GPT-42023.11 | 59.3 | 5.5 | 35.2 | 16.8 | |
| BPO + Vicuna-v1.3 13BBase LLM=Vicuna-v1.3, Model A=13B + BPO, Model B=13B, Evaluation LLM=GPT-42023.11 | 52.5 | 3.7 | 43.8 | 13.1 | |
| BPO + Llama-2-chat 7B (Cross-size)Base LLM=Llama-2-chat, Model A=7B + BPO, Model B=70B, Evaluation LLM=GPT-42023.11 | 48.8 | 3.7 | 47.5 | -7.1 |