[BFCL] Add support for LGAI-EXAONE/EXAONE-4.0 series models

[Luzzer]

Summary

This PR adds support for LG AI Research's EXAONE 4.0 series models to BFCL.

EXAONE 4.0 is a unified large language model developed by LG AI Research.
The model supports agentic tool use with Hermes-style tool calling format and provides multilingual capabilities for English, Korean, and Spanish.

The implementation follows the existing Qwen FC handler pattern, with modifications for EXAONE 4.0's specific chat template format:

• Uses [|system|], [|user|], [|assistant|], [|endofturn|] as special tokens
• Tool definitions are wrapped in <tool></tool> tags
• Tool calls are output in <tool_call>{"name": ..., "arguments": ...}</tool_call> format
• Tool responses are formatted with <tool_response></tool_response> tags

Models Added

Model	Mode	Handler
LGAI-EXAONE/EXAONE-4.0-32B	Prompting	Exaone4Handler
LGAI-EXAONE/EXAONE-4.0-1.2B	Prompting	Exaone4Handler
LGAI-EXAONE/EXAONE-4.0-32B-FC	Function Calling	Exaone4FCHandler
LGAI-EXAONE/EXAONE-4.0-1.2B-FC	Function Calling	Exaone4FCHandler

Changes

New Files

• bfcl_eval/model_handler/local_inference/exaone4.py - Prompting mode handler
• bfcl_eval/model_handler/local_inference/exaone4_fc.py - Function Calling mode handler

Modified Files

• bfcl_eval/constants/model_config.py - Added EXAONE 4.0 model configurations

EXAONE 4.0 Details

• Special Tokens: [|system|], [|user|], [|assistant|], [|endofturn|]
• Tool Call Format: Hermes-style (<tool_call>{"name": ..., "arguments": ...}</tool_call>)
• Reasoning Mode: <think>...</think> block support
• Context Length: 131,072 tokens
• Dtype: bfloat16

References

• EXAONE 4.0 HuggingFace
• EXAONE 4.0 Technical Report

Testing

Tested with:

• bfcl generate --model LGAI-EXAONE/EXAONE-4.0-32B-FC --backend vllm --num-gpus 4 --local-model-path /local/model/path --test-category multi_turn,single_turn
• bfcl evaluate --model LGAI-EXAONE/EXAONE-4.0-32B-FC --test-category multi_turn,single_turn