update hammer handler and add Hammer2.0 model

[linqq9]

Hello, we have updated the hammer handle and added Hammer2.0 series models, including Hammer2.0-7b, Hammer2.0-3b, Hammer2.0-1.5b and Hammer2.0-0.5b. The performance on BFCL-V3 is as follows:

Model	Overall Acc	Non-live AST	Non-live Exec	Live AST	Multi Turn Acc	Relevance	Irrelevance
MadeAgents/Hammer2.0-7b (FC)	56.60	90.15	82.64	68.68	15.75	92.68	68.20
MadeAgents/Hammer2.0-1.5b (FC)	51.94	84.31	81.80	63.17	11.38	92.68	61.83
MadeAgents/Hammer2.0-3b (FC)	49.88	86.77	80.25	66.06	0.50	92.68	68.59
MadeAgents/Hammer2.0-0.5b (FC)	39.51	67.00	65.73	51.62	0.00	87.80	67.00

[ShishirPatil]

Thank you for the PR @linqq9 We will review this PR tomorrow PST :)

[HuanzhiMao]

Hey @linqq9, Thanks for the PR!

Question regarding the _format_prompt function:

1. In these lines, why we remove all those <|im_start|> <|im_end|> tags? The chat template for Hammer2.0 on huggingface does include them.
2. What happened to the system prompts here? Previously they were contacted in the task instruction section, but now it seems that they are all thrown away.
3. Why do we need to special handle the situation when the length of the prompt is 2 (code here)?

[linqq9]

Hey @linqq9, Thanks for the PR!

Question regarding the _format_prompt function:

1. In these lines, why we remove all those <|im_start|> <|im_end|> tags? The chat template for Hammer2.0 on huggingface does include them.
2. What happened to the system prompts here? Previously they were contacted in the task instruction section, but now it seems that they are all thrown away.
3. Why do we need to special handle the situation when the length of the prompt is 2 (code here)?

@HuanzhiMao hi，Thanks for your questions. Here are my responses:
1·. Regarding the <|im_start|> <|im_end|> tags, in _query_prompting, we use the client.chat.completions.create function. This function already includes handling of these special symbols for the user's provided content.
2. As for the system prompts, the content in the system is later included in the generated content. Hence, there is no need to add it redundantly and it is removed.
3. We use the length of the prompt being 2 to determine if historical information is included. When the length of the prompt is 2 (system and user), it is considered that there is no historical information. If it is greater than 2, it means historical information is present.

[linqq9]

Hi @HuanzhiMao, Do you have any other questions?

[HuanzhiMao]

I will submit a PR to your branch.

[linqq9]

I will submit a PR to your branch.

ok, thanks!

[HuanzhiMao]

Hi @linqq9, I have submitted a PR to your branch. MadeAgents#2
Things I changed:

1. Let HammerHandler inherit from OSSHandler instead and copy over any necessary decoding logic from the SalesforceHandler. This simplifies things a lot.
2. Change to use completions endpoint instead of chat.completions. This won't affect the score.
3. Since Hammer doesn't take user-supplied system message, we turn any system message into user message (only message role change). This affects the live categories, and is also what we do to other models in similar situations.
4. Since Hammer has its own system prompt, we don't need to add the default BFCL system prompt in _pre_query_processing_prompting.

ps, after the change, the score is roughly the same as you reported for MadeAgents/Hammer2.0-7b (FC).

[HuanzhiMao]

LGTM

[linqq9]

Hi @linqq9, I have submitted a PR to your branch. MadeAgents#2 Things I changed:

1. Let HammerHandler inherit from OSSHandler instead and copy over any necessary decoding logic from the SalesforceHandler. This simplifies things a lot.
2. Change to use completions endpoint instead of chat.completions. This won't affect the score.
3. Since Hammer doesn't take user-supplied system message, we turn any system message into user message (only message role change). This affects the live categories, and is also what we do to other models in similar situations.
4. Since Hammer has its own system prompt, we don't need to add the default BFCL system prompt in _pre_query_processing_prompting.

ps, after the change, the score is roughly the same as you reported for MadeAgents/Hammer2.0-7b (FC).

@HuanzhiMao Thank you for your modification！