YanJo199/scientific-papers-sft-v1

Scientific Papers SFT Queue

Instruction-tuning style records derived from scientific paper passages for summarization and keyword extraction tasks.

Splits in this repo

• —pilot/sft_review_queue_10.jsonl: 10-item pilot split for quick validation
• —sft_review_queue_500.jsonl: 500-item review queue

Schema

Each JSONL row has:

• —task: summary or keywords
• —instruction: task instruction
• —input: source passage
• —draft_output: initial target output
• —review_status: review state (pending in current queue)
• —notes: free-text reviewer notes

Data quality status

• —All rows in the current uploads parse as valid JSONL
• —Known issue: some rows contain encoding artifacts (for example ’, Â)
• —Recommended next step: text normalization before production fine-tuning

Intended use

• —Pilot SFT experiments
• —Human-in-the-loop review pipelines
• —Dataset quality iteration before large-scale releases

Limitations

• —Draft outputs are not fully human-validated
• —Potential encoding artifacts can reduce model quality
• —Domain coverage depends on source paper mix

Attribution

Maintainer: YanJo199