What is the difference between these versions of OpenChat?

[RZFan525]

Thank you for your amazing work! I have some questions below:

In alpaca eval leaderboard, there are 5 versions: OpenChatV3.1, OpenChatV2-W, OpenChatV2, OpenChat, OpenChat8192. What is the difference? And what is the datasets used? What is the difference between openchat_shareGPT_v3 and openchat_shareGPT4? Which datasets do you use for OpenChatV3.1?

Looking forward to your reply.

[imoneoi]

These are different versions of the OpenChat release.

The OpenChat, OpenChat8192 were trained solely on GPT-4 answers from ShareGPT (openchat_sharegpt4_dataset).

OpenChatV3.1, OpenChatV2-W, OpenChatV2 were trained on all ShareGPT data (including GPT-3.5 and GPT-4, openchat_sharegpt_v3).

Besides, OpenChat V3 and later leveraged Llama 2 base model, while others are based on Llama 1.

[RZFan525]

These are different versions of the OpenChat release.

The OpenChat, OpenChat8192 were trained solely on GPT-4 answers from ShareGPT (openchat_sharegpt4_dataset).

OpenChatV3.1, OpenChatV2-W, OpenChatV2 were trained on all ShareGPT data (including GPT-3.5 and GPT-4, openchat_sharegpt_v3).

Besides, OpenChat V3 and later leveraged Llama 2 base model, while others are based on Llama 1.

Thank you for your detailed reply!

[RZFan525]

These are different versions of the OpenChat release.

The OpenChat, OpenChat8192 were trained solely on GPT-4 answers from ShareGPT (openchat_sharegpt4_dataset).

OpenChatV3.1, OpenChatV2-W, OpenChatV2 were trained on all ShareGPT data (including GPT-3.5 and GPT-4, openchat_sharegpt_v3).

Besides, OpenChat V3 and later leveraged Llama 2 base model, while others are based on Llama 1.

In addition to the datasets used, what did you update in OpenChatV2-W comparing to OpenchatV2?

[RZFan525]

Sorry to trouble you again.

I'm wonder what is updated in OpenChatV3.2 SUPER which boosts the score of MT-Bench.

[imoneoi]

Sorry to trouble you again.

I'm wonder what is updated in OpenChatV3.2 SUPER which boosts the score of MT-Bench.

Basically hyperparameter tuning :(

[imoneoi]

These are different versions of the OpenChat release. The OpenChat, OpenChat8192 were trained solely on GPT-4 answers from ShareGPT (openchat_sharegpt4_dataset). OpenChatV3.1, OpenChatV2-W, OpenChatV2 were trained on all ShareGPT data (including GPT-3.5 and GPT-4, openchat_sharegpt_v3). Besides, OpenChat V3 and later leveraged Llama 2 base model, while others are based on Llama 1.

In addition to the datasets used, what did you update in OpenChatV2-W comparing to OpenchatV2?

V2-W leveraged weighted+conditioned finetuning, while V2 is conditioned finetuning

[RZFan525]

These are different versions of the OpenChat release. The OpenChat, OpenChat8192 were trained solely on GPT-4 answers from ShareGPT (openchat_sharegpt4_dataset). OpenChatV3.1, OpenChatV2-W, OpenChatV2 were trained on all ShareGPT data (including GPT-3.5 and GPT-4, openchat_sharegpt_v3). Besides, OpenChat V3 and later leveraged Llama 2 base model, while others are based on Llama 1.

In addition to the datasets used, what did you update in OpenChatV2-W comparing to OpenchatV2?

V2-W leveraged weighted+conditioned finetuning, while V2 is conditioned finetuning

Many many thanks for your reply!

Have a good time!

[hy-2013]

These are different versions of the OpenChat release.

The OpenChat, OpenChat8192 were trained solely on GPT-4 answers from ShareGPT (openchat_sharegpt4_dataset).

OpenChatV3.1, OpenChatV2-W, OpenChatV2 were trained on all ShareGPT data (including GPT-3.5 and GPT-4, openchat_sharegpt_v3).

Besides, OpenChat V3 and later leveraged Llama 2 base model, while others are based on Llama 1.

openchat 3.5 is based on Mistral-7B? Based on llama1/2 or Mistral without adding glossaries and post-pretrain, can such good results be achieved only with SFT and C-RLFT? thanks