tablinum/cowork-companion-sft

T

Tablinum

cowork-companion-sft

the 3,258 rows that turn a stock base model into an office coworker

Oracle-verified SFT corpus — synthetic, with attributed sources · the exact mix used to train [cowork-companion](https://huggingface.co/tablinum/cowork-companion) (354M) and [cowork-companion-1.2b](https://huggingface.co/tablinum/cowork-companion-1.2b)

What this is

The complete supervised-fine-tuning corpus behind both Cowork Companion models (the "v10 mix"). Stock bases are near-chance on office-production tasks (≈0.24); trained on these 3,258 rows, a 354M model reaches 1.000 on the frozen instruments. 2,500 rows are program-generated synthetic (2,100 capability + 400 replay); 658 are teacher-distilled traces and 100 derive from 10 sourced third-party facts — both attributed below. No external documents, no user data.

stream	rows	what it teaches
synthetic capability rows	2,100	seven families: pptx_create, xlsx_build, docx_edit, doc_ingest, cross_doc_conflict, grounded_refusal, table_aggregation — with the magnitude curriculum, real-shaped variants, and instruction-omission coins
oracle-filtered teacher traces	658	598 docx edits + 60 xlsx sheet_specs distilled from LFM2.5-2.6B; every trace oracle-verified (658 of 1,250 drawn kept)
replay rows	400	stability replay from the earlier curriculum
2-shot selector rows	100	format-stability selector demos; their 10 underlying facts are sourced from Qwen/RationaleRM (CC-BY-4.0), curated and normalized — see Attribution

Row format

JSONL; one chat per row:

{"id": "cc-t8-pptx-001", "stream": "pptx_create",
 "messages": [{"role": "system", "content": "...coworker contract..."},
              {"role": "user", "content": "<instruction + document text>"},
              {"role": "assistant", "content": "{\"answer\": ..., \"artifact\": {...}, ...}"}]}

The system prompt is the model's contract (single JSON envelope: answer / artifact / claims / overall_confidence / caveats); the assistant targets are envelope JSON — the same shape the trained models emit.

The design rules baked in

• Grounded refusal is a trained skill — unanswerable items teach artifact: null and "not specified in the material", never invented figures.
• Numbers only from the material — every numeric claim carries a provenance discipline; multi-row arithmetic is deliberately NOT taught in-model (it fails at every scale; route it to a harness).
• Curriculum honesty — the aggregation stream carries a magnitude curriculum (small tables dominate) reflecting the measured 3-row boundary at 354M.

Attribution — third-party content in the mix

• Teacher traces (658 rows). These are outputs of LiquidAI LFM2.5-2.6B (oracle-filtered), which makes this corpus a derivative work of the teacher. The teacher is distributed under the LFM Open License v1.0; its conditions — including the commercial-use revenue threshold — propagate to these rows and to models trained on them.
• Selector rows (100 rows). Derived from 10 technical facts sourced from the human-checklist split of Qwen/RationaleRM, used under CC-BY-4.0. Content and register come from the source; selection was curated and facts normalized per row. Attribution notice: Contains material derived from Qwen/RationaleRM (Qwen team), licensed CC-BY-4.0.
• No other third-party content. The remaining 2,500 rows are program-generated by this project's own deterministic generators.

License — read before commercial use

Free for individuals, non-commercial use, and organizations under US$1M annual revenue — attribution required ("Training data: cowork-companion-sft by Tablinum / Chad St. Thomas"). Organizations at or above US$1M annual revenue need a commercial license. Full terms: DATASET_LICENSE.txt.

Provenance

Built by deterministic, seed-pinned generators in the CuriousGene repository; sha256 of train.jsonl: dd2ac333dcf2d2b1…. The eval instruments used to verify models trained on this corpus are deliberately NOT published, to keep them uncontaminated — their hashes are published (INSTRUMENTS.sha256) so every published number is traceable to the exact frozen instrument. The deployment harness for the trained models is open source: github.com/chadstthomas-svg/cowork-cli.