Kiria-Nozan/TRIM-gpt-oss-120b-16-tasks-wo-playbook

---
configs:
- config_name: default
data_files:
- split: train
path: "train/*.jsonl"
---

# TRIM Agent Reasoning Messages Without Playbook (HF Public Export)

This directory is a Hugging Face-friendly public export of the TRIM agent reasoning SFT data.
This variant was built from the no-playbook rewrite outputs generated with `vllm/gpt-oss-120b`.

## What Is Included

- Provider: `vllm`
- Model: `gpt-oss-120b`
- Splits present: `train`
- Records in this export manifest: `14006`
- Tasks in this split: `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, hERG`

## Record Schema

Each JSONL line is one training example with these top-level fields:

- `schema_version`
- `task`
- `split`
- `sample_index`
- `sample_id`
- `smiles`
- `gt_label`
- `final_answer_option`
- `messages`

The `messages` field stores a tool-augmented chat transcript, including nested `tool_calls` and the assistant `thinking` text used in the original SFT export.

## Public Sanitization

- Local absolute `source_paths` have been removed from the sample records by default.
- Task-level export metadata is stored under `metadata/manifest.json`.

## Loading Example

```python
from datasets import load_dataset

ds = load_dataset(
"json",
data_files={"train": "train/*.jsonl"},
)
```