i can not find tool_llama model,can you provide the link?thank you about toolbench HOT 10 CLOSED

Hi, we have released the lora version here. Tool-llama trained with full parameters will be released soon. Please keep following our updates.

from toolbench.

Thanks for your reply, and the another i want to ask you is "can the weight i fine_tune to replace the "lora_version toolllama weight " as detal weight "

from toolbench.

If you train a lora version of tool llama model, sure you can replace the released lora weights with your own weights. If you train with full parameters, you can replace the /path/to/ToolLLaMA/weights with your saved model path:
python toolbench/inference/inference_single_tool.py --tool_name weather --model_path /path/to/ToolLLaMA/weights

from toolbench.

thank you! Before infernece(i have finetuned a weight), should i perform this step(apply base weight and delta weight) to get tool_llama weight to inference? And i can not see the delta weight, the link you shared is data not a model.

from toolbench.

No, you do not need to apply any delta weight if you fine-tune a model yourself. Just replace the /path/to/ToolLLaMA/weights with your own model path~

from toolbench.

thank you for your reply

from toolbench.

Dear, the multi data, such as meta_file.json and the first conversation of it, some action look like not a tool use,i want to know what the mean and purpose(i guess that is just to teach LLM how to do, but i don know when inference LLM how to achieve that)

from toolbench.

Detailed information about the tools can be found in BMTools. In multi tools scenario, we select several tools from BMTools.

Specifically, in meta analysis scenario, chemical-prop/meta_analysis/Slides Making/Wikipedia/file_operation/Bing_search are selected. The actions shown above are from Slides Making.

from toolbench.

The weights saved bad(can not be loaded),which show cuda out of memory,how to do that?

from toolbench.

error occur when training : Failed to clone() tensor with name _fsdp_wrapped_module.model.layers.31.mlp.up_proj.weight on rank 0. This may mean that this state_dict entry could point to invalid memory regions after returning from state_dict() call if this parameter is managed by FSDP,

from toolbench.