book-sft-pipeline
muratcankoylan/Agent-Skills-for-Context-Engineering / examples/book-sft-pipeline
guides converting a book into an SFT dataset , which improves: faithful imitation of an author's voice
Details
- judge’s score
- 7/10
- place
- outside the packs
- publisher
- muratcankoylan
- status
- new
- pinned commit
- 9ff0c99 on GitHub (new tab)
- skill’s last commit
- Jun 29, 2026
- repo stars
- 18,177
- main type
- Other
Judgement
Precise trigger and clear scope, but outside Claude Code. Code often pseudo-code and inconsistent (chunk sizes, learning rate, await). Annexes truncated. Score 7 out of 10.
reservations
- Truncated annexes: references/segmentation-strategies.md, tinker-format.md and tinker.txt were only partly read, and scripts/pipeline_example.py is not provided.
- Internal inconsistencies: chunks of 150-400 words in SKILL.md versus 250-650 in the annex, learning rate 5e-4 versus 1e-4 x2, and non-executable examples (await in non-async code, undefined chunk.id).
- Result and cost figures presented without a verifiable protocol (Pangram, 70% human, ~$2), so not reproducible.
- The skill does not address copyright: training a model on entire books to imitate a living author raises a legal question absent from the text.
- Triggering
- 2/2
- Added value
- 1/2
- Precision
- 1/2
- Scope
- 2/2
- Quality
- 1/2
Alerts
2 alerts, by category
Network
- high risk
The pipeline sends full book texts to third-party APIs (low-cost LLM for instructions, Tinker platform) with no warning about confidentiality or rights.
Static scan
- info risk
bundled scripts, not run by the SKILL.md
scripts/pipeline_example.py
Original description
This skill should be used for book-to-SFT pipelines: ePub extraction, literary segmentation, author-voice dataset construction, style-transfer training, LoRA workflows, and model evaluation for voice replication.
added on Oct 19, 2026 · license MIT