README.md

---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: task_type
    dtype: string
  - name: task_config
    dtype: string
  - name: casetype
    dtype: string
  - name: facts
    dtype: string
  - name: instruction
    dtype: string
  - name: ground_truth_text
    dtype: string
  - name: ground_truth_items
    sequence: string
  - name: ground_truth_raw_json
    dtype: string
  - name: legal_category
    dtype: string
  - name: legal_category_name
    dtype: string
  - name: legal_category_confidence
    dtype: float32
  - name: legal_category_parse_status
    dtype: string
  - name: legal_category_rationale
    dtype: string
  - name: legal_category_tag_model
    dtype: string
  - name: legal_category_tag_source
    dtype: string
  - name: domain
    dtype: string
  - name: dataset
    dtype: string
  - name: scoring_kind
    dtype: string
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: valid
    path: data/valid-*
  - split: test
    path: data/test-*
---

# LBox Open classification formatted for MAE / SFT evaluation

This dataset is built from local `export/lbox` JSONL files.

- `id` is preserved for joining with inference outputs.
- `instruction` is the exact prompt body used by this repo.
- `ground_truth_items` stores one or more labels with a stable list schema.
- `ground_truth_text` is a display-friendly joined label string.
- `ground_truth_raw_json` preserves the original local label object.
- `legal_category*` columns are LLM-generated legal-topic tags produced with
  `google/gemma-4-26B-A4B-it`. The current taxonomy merges the earlier
  `family_case` and `patent_ip` tags into `family_patent_special`.