Kiria-Nozan/TRIM-gpt-oss-120b-separate-neighbors-only-4nbr

TRIM Agent Reasoning Messages（HF公共导出）是一个Hugging Face友好的公共导出数据集，包含TRIM代理推理的监督微调（SFT）数据。数据由vllm提供，基于gpt-oss-120b模型，采用local_neighbor_only的SFT模式。数据集仅包含训练集分割，共有11727条记录，覆盖多个任务，如AMES、BBB_Martins、Bioavailability_Ma、CYP2C9_Substrate_CarbonMangels、CYP2D6_Substrate_CarbonMangels、CYP3A4_Substrate_CarbonMangels、Carcinogens_Lagunin、ClinTox、DILI、HIA_Hou、PAMPA_NCATS、Pgp_Broccatelli、SARSCoV2_3CLPro_Diamond、SARSCoV2_Vitro_Touret、Skin_Reaction和hERG。每条记录以JSONL格式存储，包含schema_version、sft_mode、task、split、sample_index、sample_id、smiles、gt_label、final_answer_option和messages等字段，其中messages字段存储工具增强的聊天记录，包括嵌套的tool_calls和助理思考文本。数据经过公共清理，移除了本地绝对路径，任务级导出元数据存储在metadata/manifest.json中。TRIM Agent Reasoning Messages (HF Public Export) is a Hugging Face-friendly public export of the TRIM agent reasoning supervised fine-tuning (SFT) data. The data is provided by vllm, based on the gpt-oss-120b model, using the local_neighbor_only SFT mode. The dataset includes only the train split with 11727 records, covering multiple tasks such as AMES, BBB_Martins, Bioavailability_Ma, CYP2C9_Substrate_CarbonMangels, CYP2D6_Substrate_CarbonMangels, CYP3A4_Substrate_CarbonMangels, Carcinogens_Lagunin, ClinTox, DILI, HIA_Hou, PAMPA_NCATS, Pgp_Broccatelli, SARSCoV2_3CLPro_Diamond, SARSCoV2_Vitro_Touret, Skin_Reaction, and hERG. Each record is stored in JSONL format with fields including schema_version, sft_mode, task, split, sample_index, sample_id, smiles, gt_label, final_answer_option, and messages, where the messages field stores a tool-augmented chat transcript, including nested tool_calls and assistant thinking text. The data has been publicly sanitized by removing local absolute source paths, and task-level export metadata is stored under metadata/manifest.json.