Is the conversation template tied to a `model_id` or to a `model_path`?

Nobody has claimed this yet.

In `gen_model_answer.py`, a conversation template is loaded by checking `model_id`

https://github.com/lm-sys/FastChat/blob/8163cb2719b3155fd5b83dd0bf4190f61a847a6d/fastchat/llm_judge/gen_model_answer.py#L107

However, the function definition suggests to me that `model_path` might be the intended input:

https://github.com/lm-sys/FastChat/blob/8163cb2719b3155fd5b83dd0bf4190f61a847a6d/fastchat/model/model_adapter.py#L374-L377

Other parts of the codebase seem to suggest this as well, e.g., `load_model` seems to `load_adapter(model_path)`

https://github.com/lm-sys/FastChat/blob/8163cb2719b3155fd5b83dd0bf4190f61a847a6d/fastchat/model/model_adapter.py#L196

If `model_id == model_path` (or at least they are somewhat similar), then I think this doesn't cause a problem (e.g., if model path points to an HF repo and the model_id is also the model's name, e.g., model_path is `databricks/dolly-v2-12b` and model_id is `dolly-v2`), but if they are wildly different (e.g., maybe model_id is an internal name given to the model) then this can lead to the wrong conversation template getting loaded.

I _suspect_ that this might be a slight bug, and `model_path` should be the one used in

https://github.com/lm-sys/FastChat/blob/8163cb2719b3155fd5b83dd0bf4190f61a847a6d/fastchat/llm_judge/gen_model_answer.py#L107

...but maybe this behavior is actually expected. If it is expected, I think this should be clarified in the [docs over here](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge#evaluate-a-model-on-mt-bench) since (at least to me) the docs suggest that the model id is just a label, e.g., for the benchmark itself.

Either way, thanks for making this library---it's super useful and very easy to use! 😃

Read fastchat/llm_judge/gen_model_answer.py around line 107 and fastchat/model/model_adapter.py around lines 196 and 374-377 to trace whether conversation templates use model_id or model_path. Check the MT-Bench evaluation documentation, then establish the intended identifier and either correct the lookup or clarify the behavior in the docs.

Written by the indexing model from the issue text.