khairi/kothar-dataset-it

Kothar fine-tuning datasets

This repo documents the instruction-tuning datasets built by this repository for a protein language model. All datasets are seeded from the Neo4j protein knowledge graph (docs/neo4j_schema.md) or from computed sequence features, built by the scripts under scripts/python/, and published as HuggingFace Hub configs at khairi/kothar-dataset-it.

What's here

Document	Covers
shared-conventions.md	Sequence encoding, the computed sequence-feature set, and the reasoning-trace triple format used by every dataset below
swissprot-split.md	The full-corpus SwissProt identity split that some pipelines build on
annotation.md	Protein → molecular function / biological process / cellular component
scop.md	Protein → SCOP structural classification (fold / superfamily / family)
generation.md	SCOP fold / family name → matching protein sequence (the reverse direction of scop.md)
ppi.md	Protein pair → do they interact (yes/no)

The eight dataset configs

Each pipeline produces a thinking variant (with a <think>...</think> reasoning trace before the answer) and a non-thinking variant (direct instruction → answer, used as a direct-SFT ablation baseline). All eight are published as separate configs in the same Hub repo, each with train / validation / test splits:

Config	Train	Validation	Test	Row shape
annotation_thinking	1,057,896	40,185	12,870	Entry, TargetRelation, Instruction, Answer, Reasoning
annotation_non_thinking	1,057,896	40,185	12,870	Entry, TargetRelation, Instruction, Answer
scop_thinking	38,999	3,235	3,081	Entry, TargetRelation, Instruction, Answer, Reasoning
scop_non_thinking	38,999	3,235	3,081	Entry, TargetRelation, Instruction, Answer
generation_thinking	27,510	2,213	1,755	Entry, ConditionRelation, ConditionValue, Instruction, Answer, Reasoning
generation_non_thinking	31,237	2,537	2,483	Entry, ConditionRelation, ConditionValue, Instruction, Answer
ppi_thinking	121,150	772	2,334	EntryA, EntryB, Label, Instruction, Answer, Reasoning
ppi_non_thinking	121,150	772	2,334	EntryA, EntryB, Label, Instruction, Answer

generation_thinking has fewer rows than generation_non_thinking — see generation.md for why. Every other thinking/non-thinking pair is row-for-row identical apart from the Reasoning column.

Common pipeline shape

Every dataset (except ppi, which reverses steps 1 and 2 — see ppi.md) is built in the same three stages, run from the project root:

1. build_pairs.py — query the graph for the full eligible pool, build the instruction/answer text (and, for pipelines whose thinking trace doesn't need a separate pass, the reasoning trace too), write instructions/<name>/pairs.jsonl + rejected.jsonl.
2. split_by_identity.py — MMseqs2 sequence-identity split (validation ≤70% identity to any train entry, test ≤30%) over the entries in pairs.jsonl, so validation/test measure generalization to remote homologs rather than near-duplicates of training sequences. Output: instructions/<name>/splits/<name>_split.csv.
3. build_datasets.py — join pairs with their split, assemble the _thinking / _non_thinking HuggingFace DatasetDict configs, write local Parquet (--save-dir data/finetune) and/or push to the Hub (--repo-id ... --push).

All four pipelines share scripts/python/shared/features.py for computed sequence features — see shared-conventions.md.

Design principle: no future-relation leakage

Every dataset here answers the question "would a biologist know this from a genuinely novel, uncharacterized sequence?" Context/reasoning content that requires prior experimental characterization (curated GO/EC/UniProt annotations, homology-transferred family names, etc.) is excluded even when it correlates strongly with the target — see each dataset's own doc for the specific redundancy checks that drove those exclusions.