Training data
FinFlow creates training files. Your downstream program trains the target model. Explicit DoReMi/RegMix commands also train small proxies to learn mixture weights.
Daily v7 flywheel outputs
Original text, visual descriptions, and optional translations/rewrites pass independent review and are segmented with the target tokenizer. Related content shares a train/validation split; cross-day deduplication preserves all origins.
See configuration, source admission, and releases and snapshots.
Existing v5/v6 financial releases
Use training.cli to export body-text corpora from existing packages; see Existing release export. This entry point does not call translation or rewriting models and does not use the flywheel’s target-tokenizer recipe.
Supervised fine-tuning data
Approved v7 original text and tables can also generate document QA, structured extraction and table-calculation examples. Outputs retain complete messages, evidence, review and lineage without CPT-style text segmentation. See SFT data generation for the standalone CLI, daily integration, resumable caches and split limitations.
Visual SFT
Optional config/sft-vision.json supports visual QA and table-image-to-JSON transcription. Exports package original images, typed messages and visual evidence metadata, with separate text/vision training files. See Visual SFT.
Methods and mixtures
- Generation methods: CPT knowledge transformations, multi-step SFT and visual conversations.
- Data mixtures and selection: fixed/temperature quotas, family caps, quality/diversity selection and actual proportions.
- DoReMi / RegMix: domain preparation, real local proxy training, resume and learned weights.