Training-data generation methods

Tasks define the skill, strategies define candidate generation, and mixture recipes select approved examples. Existing recipes default to direct. New methods require explicit configuration.

CPT transformations

Add any of these names to methods in config/flywheel.local.json:

Method Output
distill Information-dense prose retaining necessary qualifications
textbook Educational organization of concepts and rules explicitly explained by the source
knowledge_list Facts retaining subjects, metrics, values, units and periods
diverse_qa Diverse reading-comprehension QA text supported by the document

Configure the shared FIN_DOC_PRETRAIN_SYNTHESIZE_API_URL, _API_KEY, _MODEL role in .env. Independent review continues to use FIN_DOC_PRETRAIN_REVIEW_*; translate/rewrite retain their existing roles. Endpoints are OpenAI-compatible base URLs. Candidates are stored separately from original text and must not add outside facts.

These transformations adapt Nemotron-CC task patterns to FinFlow evidence, review and lineage, without installing NeMo. CPT QA is plain training text; it is not automatically approved SFT.

Multi-step text SFT 

cp config/sft-methods.json config/sft-methods.local.json
python -m training.sft_cli --config config/sft-methods.local.json \
  --flywheel-config config/flywheel.local.json preflight
python -m training.sft_cli --config config/sft-methods.local.json \
  --flywheel-config config/flywheel.local.json \
  build --dataset-id sft-methods-001 --release YOUR_RELEASE_ID

Configure tasks and strategies separately. Non-direct strategies currently support document_qa only. Extraction, verified calculations and transcription retain specialized direct contracts. Every enabled task/strategy must have a compatible counterpart, or preflight fails.

Strategy Implemented stages Adaptation boundary
direct Generate → deterministic checks → independent review Existing baseline
self_instruct Expand curated seeds → screen instructions → generate answers → review One source-bound expansion; no cross-document self-bootstrapping seed pool
answer_first Select verbatim answer → inverse question → screen → review original answer API adaptation; no reverse-model training or iterative Humpback fine-tuning
evol_instruct Basic questions → constraints/scope/comparison evolution → information-gain and answerability screening → answers/review Confined to current evidence; no open-domain topic expansion
codeclm Encode use case/skills/rubric → decode instructions → rubric refinement → answers/review Optional target-model comparison; no target-model fine-tuning

Sources: Self-Instruct, Evol-Instruct, Instruction Backtranslation, CodecLM.

Configuration

  • seed_instructions: optional 1–100 seed instructions, up to 2000 characters each; five task shapes are built in. Do not put validation answers in the seed library.
  • strategy_seed: seed/operator selection seed, default 42.
  • evolution_rounds: 1–5, default 2.
  • contrastive_filter: false by default; true requires codeclm and FIN_DOC_SFT_TARGET_*.
  • contrastive_min_gap: integer teacher-minus-target rubric score gap, 1–5, default 1.

Generation uses FIN_DOC_SFT_GENERATE_*; instruction screening and final review use FIN_DOC_SFT_REVIEW_*. For contrastive filtering, the target receives source evidence and the question. The reviewer scores both answers against the rubric and uses the score gap to select candidates for final factual review.

Jobs retain intermediate responses, stage artifacts, model roles and final review. Samples expose strategy and strategy_stages. Recipe identities include strategy/seed settings and model identities; the engine is now sft-v3, while export contracts remain SFT v2. Existing packages remain immutable. Multi-step methods consume more requests; after a partial run, use a new dataset ID with the same recipe to reuse cached responses. Do not concatenate cumulative resume snapshots. Seeds stabilize workflow choices, not remote model outputs; cached responses preserve what actually happened.

Additional visual tasks

Copy config/sft-vision-methods.json to a local recipe to enable:

  • visual_description: visible structure, labels, units and trends; no guessed precise values from ambiguous charts.
  • visual_conversation: 2–max_conversation_turns coherent QA turns about the same image; default maximum 4, supported 2–8.

Use the existing FIN_DOC_SFT_VISION_GENERATE_* and FIN_DOC_SFT_VISION_REVIEW_* roles. Review checks each turn against the original image and rejects contradictions or unsupported inferences. Only the first user message includes the image placeholder. Per-turn limits and the total max_answer_chars cap apply. Existing visual QA and table JSON transcription can run alongside them. Category design follows LLaVA’s data guide; using actual vision APIs is an adaptation of that workflow.

Selection and publication

Generation counts do not determine final training proportions. Review and deduplicate before mixture selection. DoReMi / RegMix run explicit local proxy training independently of daily SFT/API calls. training/method_cards.py records attribution and adaptation boundaries; relevant SFT/experiment manifests include these cards.