yamaTK/merged_dataset_final_clean_v41
merged_dataset_final_clean_v41 English Rule-based cleaned SFT dataset for structured output generation (JSON / YAML / XML / TOML / CSV). Data Source This dataset was built from competition-provided datasets only. The cleaning pipeline loads the following source groups: u-10bei (6 datasets: source ids 1-1 to 1-6) daichira (3 datasets: source ids 2-1 to 2-3) After strict filtering and sampling for v4.1, the final retained rows are from u-10bei… See the full description on the dataset page: https://huggingface.co/datasets/yamaTK/merged_dataset_final_clean_v41.
mergeddatasetfinalcleanv41
English
Rule-based cleaned SFT dataset for structured output generation (JSON / YAML / XML / TOML / CSV).
Data Source
This dataset was built from competition-provided datasets only.
The cleaning pipeline loads the following source groups:
- u-10bei (6 datasets: source ids
1-1to1-6) - daichira (3 datasets: source ids
2-1to2-3)
After strict filtering and sampling for v4.1, the final retained rows are from u-10bei sources (1-1 to 1-6).
Processing (What was done to original data)
merged_dataset_final_clean_v41.jsonl is generated by deterministic, rule-based preprocessing (no LLM rewriting).
Main steps:
- Merge and normalize records into a unified message format (
system,user,assistant). - Parse/validate assistant outputs by target format (CSV/JSON/YAML/XML/TOML) with rule checks.
- Remove duplicates (exact hash + near-duplicate handling in the cleaning pipeline).
- Remove extra non-structured text and malformed fenced outputs.
- Apply consistency and outlier filtering.
- Remove
Approach: ... Output:style CoT prefix and keep only final output payload where applicable. - Apply aggressive, deterministic sampling to match training constraints and budget.
Mechanical-Only Transformation Statement
- No LLM was used to generate, rewrite, paraphrase, or synthesize training samples.
- Transformations are implemented with deterministic Python logic (parsers, validators, hashing, filtering, sampling).
public_150.jsonwas not modified during dataset construction.
Dataset Statistics (v41 file)
- File:
merged_dataset_final_clean_v41.jsonl - Total rows:
1000 - Target format distribution:
json:332yaml:234toml:164xml:135csv:135- Source ids in final file:
1-1,1-2,1-3,1-4,1-5,1-6
Columns
Each JSONL row contains:
sourceseriesmessages(chat-style list with roles/content)formatsource_formatcomplexityschematypen_tokens
日本語
構造化出力(JSON / YAML / XML / TOML / CSV)向けSFT用に、ルールベースでクリーニングしたデータセットです。
元データ
コンペ提供データのみを使用しています。
クリーニングパイプラインでは次の系列を読み込みます。
- u-10bei(6データセット:
1-1〜1-6) - daichira(3データセット:
2-1〜2-3)
ただし v4.1 の厳しめのフィルタとサンプリング後、最終的に残った行は u-10bei 系(1-1〜1-6)です。
元データに対して行った処理
merged_dataset_final_clean_v41.jsonl は、決定的なルール処理のみで生成しています(LLMによる書き換えなし)。
主な処理:
- レコードを
system / user / assistantの統一形式に正規化 - 出力フォーマット別(CSV/JSON/YAML/XML/TOML)に構文検証
- 重複除去(完全一致ハッシュ + 近似重複処理)
- 余計な説明文や不正なフェンス付き出力の除去
- 整合性チェックと外れ値フィルタ
Approach: ... Output:形式のCoT接頭部を除去し、最終出力本体のみ保持- 学習制約に合わせた決定的サンプリング
機械的変換のみであること
- 学習データの生成・書き換え・言い換え・合成にLLMは使っていません。
- 変換はすべて、Pythonによる決定的な処理(パース・検証・ハッシュ・フィルタ・サンプリング)です。
- データ構築時に
public_150.jsonは変更していません。
データ統計(v41)
- ファイル:
merged_dataset_final_clean_v41.jsonl - 総件数:
1000 - 目標フォーマット内訳:
json:332yaml:234toml:164xml:135csv:135- 最終ファイル内の source id:
1-1,1-2,1-3,1-4,1-5,1-6
カラム
各JSONL行の主なキー:
sourceseriesmessagesformatsource_formatcomplexityschematypen_tokens
