annotation-training-csv

name	annotation-training-csv
description	Convert SuperLeaf annotation training exports, records.jsonl files, or export ZIPs into compact CSV training datasets. Use this whenever the user mentions annotation-training-export, records.jsonl, 批注训练数据, clean training data, CSV dataset, reducing token cost, or turning annotation evaluations into LLM fine-tuning/evaluation rows.
nav_exclude	true

Annotation Training CSV

Use this skill to turn SuperLeaf annotation training exports into a clean, small CSV file for downstream training, evaluation, or manual review.

The goal is token economy: keep only the fields needed for the next training step, and avoid carrying hashes, ranges, full threads, tags, or document metadata unless the user explicitly asks for audit columns.

Default Output

Always default to this CSV schema:

id,comment,source_text,label

• id: integer starting from 0, assigned during extraction.
• comment: the annotation or review comment.
• source_text: the original source text. Prefer context.current_line_content; fall back to the line containing the range in old exports; finally fall back to target_text.
• label: the user verdict, normally positive or negative.

Do not include surrounding_before, surrounding_after, current_document_content, raw thread messages, document hashes, or ranges in the default CSV. Those fields are useful for audit/debug, but they increase token cost and may leak more context than needed.

Deterministic Script

Prefer the bundled script instead of asking an LLM to parse JSON manually:

python3 docs/skills/annotation-training-csv/scripts/extract_training_csv.py \
  "/path/to/annotation-training-export.zip" \
  training_data.csv \
  --only-training-candidates

The script accepts any of these inputs:

• an export ZIP containing records.jsonl
• an extracted export directory containing records.jsonl
• a direct records.jsonl path

Use --include-meta only when the user needs traceability columns for audit.

Skill Download

To copy this project-local skill into another workspace with npm tooling:

npm exec --yes degit OhMyYuwan/SuperLeaf/docs/skills/annotation-training-csv ./annotation-training-csv

If the repository is private, clone it with GitHub authentication first and copy docs/skills/annotation-training-csv/ manually.

Workflow

1. Locate the export ZIP, export directory, or records.jsonl.
2. Run the bundled script to produce CSV.
3. Inspect the first few rows and confirm the columns are minimal.
4. If the user needs only selected training data, rerun with --only-training-candidates.
5. If the user needs auditability, rerun with --include-meta, but keep that as a separate file from the token-efficient training CSV.

Label Rules

Use evaluation.verdict as the label. Preserve the original vocabulary instead of inventing new labels:

• positive: useful annotation or review
• negative: not useful annotation or review

When a record has no verdict, leave label empty rather than guessing.