Wombat-7B，Wombat-7B-gpt4 and ChatGPT Results on Comparison based on Vicuna test set, evaluation by gpt-4.

[onlyfish79]

1. Wombat-7B and ChatGPT Comparison based on Vicuna test set, score by GPT-4 Evaluation.

Wombat-7B: 599.0  average score: 7.5
ChatGPT: 710.5    average score: 8.9
wombat-7b / gpt35 = 84.31%

2. Wombat-7B-gpt4 and ChatGPT Comparison based on Vicuna test set, score by GPT-4 Evaluation.

Wombat-7B-gpt4: 577.0  average score: 7.2
ChatGPT: 734.5         average score: 9.2
wombat-7b-gpt4 / gpt35 = 78.13%

Wombat-7B and Wombat-7B-gpt4: use the script recover_wombat_7b.sh

According to the above results, Wombat-7B has better results than Wombat-7B-gpt4, does the result meet expectations?

[GanjinZero]

Yes, it does meet our expectations, and we observe a similar score in Wombat-7B-gpt4 vs ChatGPT. The reason is Wombat-7B uses 5 responses for one query to train RRHF. Although Wombat-7B-gpt4 uses better responses, but it only contain 2 responses for one query. We think more diverse responses are the most important point of training RRHF.

[GanjinZero]

Another possible thing is Wombat-7B use responses from its initial checkpoint, while Wombat-7B-gpt4 does not use the response from its initial checkpoint. If RRHF is trying to improve based on itself, not using responses from its initial checkpoint worse RRHF's performance.

[onlyfish79]

Understood, thank you for your response. May I ask about the upcoming roadmap for RRHF?

[GanjinZero]

Understood, thank you for your response. May I ask about the upcoming roadmap for RRHF?

Chain of thought reasoning & Scaling to 13b, 30b, 65b llama / alpaca