README.md

---
dataset_info:
  features:
  - name: grade
    dtype: int64
  - name: question
    dtype: string
  - name: golden
    dtype: string
  - name: reasoning_step
    dtype: int64
  - name: num_digits
    dtype: int64
  - name: prompt_id
    dtype: string
  - name: trans_resoning
    dtype: string
  - name: translated_question
    dtype: string
  - name: trans_model
    dtype: string
  - name: trans_timestamp
    dtype: string
  - name: trans_stop_reason
    dtype: string
  - name: COT_answer
    dtype: string
  - name: COT_model
    dtype: string
  - name: COT_timestamp
    dtype: string
  - name: COT_stop_reason
    dtype: string
  - name: source
    dtype: string
  - name: think
    dtype: string
  - name: translated_think
    dtype: string
  - name: final_answer
    dtype: string
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: simple_messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  splits:
  - name: train
    num_bytes: 20221478.205221616
    num_examples: 1627
  download_size: 9237020
  dataset_size: 20221478.205221616
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# cmath_zhtw

## 污染修復情況（2026-08-11）

- **修復方式**：污染掃描（詞庫驅動，`/data/sdg_data/sdg_pipeline/lexicon/lexicon.yaml`）→ scan-triage 逐筆人工判讀「真污染 / 誤報」→ 只修真污染；簡繁/大陸用語為**逐詞判讀替換，非 OpenCC 整篇轉換**（避免把已正確的繁體改壞）
- **修復範圍**：僅修改 `assistant` 角色內容（`user`/`tool`/`system` 未動，訓練時本就被 mask）
- **本次修復筆數**：污染清理 25 筆 + 「方程→方程式」用語修正 128 處
- **完整逐筆 before/after 對照**：見 `cleanup_0811_before_after.jsonl`（污染清理）、`fangcheng_fix_0811.jsonl`（方程→方程式）