Data mixtures and selection

training.mixture_cli consumes immutable CPT v2 or SFT v2 packages, preserves lineage and publishes a new training mixture. It makes no model calls. CPT and SFT use separate recipes. Original validation records remain intact and are not selected by training quotas.

Preview and publish

cp config/mixture-cpt.json config/mixture-cpt.local.json
# Edit inputs, grouping and budget. Relative paths are relative to the working directory (repository root).
python -m training.mixture_cli --config config/mixture-cpt.local.json plan
python -m training.mixture_cli --config config/mixture-cpt.local.json \
  build --output data/training/mixtures/cpt-001
python -m training.mixture_cli verify data/training/mixtures/cpt-001

For CPT, prefer a complete cumulative snapshot from workflow.flywheel_cli snapshot. Missing historical samples referenced by selected deltas cause an error. For SFT, start with config/mixture-sft.json; text and vision packages may be combined. Overlapping snapshots are deduplicated by content while preserving origins. Content, images or known document families crossing splits cause an error.

Parameters

Setting	Meaning
`group_fields`	One or more of `method`, `task`, `strategy`, `source_name`, `language`, `modality`
`method`	`fixed`, `temperature`, `doremi`, `regmix`
`unit`	CPT: `tokens` or `samples`; SFT: `samples`
`budget`	Positive total training quota
`max_per_family`	Optional maximum selected sample count per known document family, including derivatives
`seed`	Stable group and within-group ordering seed
`redistribute`	Reallocate unused quota to nonzero-weight groups with remaining capacity; default false
`allow_shortfall`	Allow publication below budget; default false

Multiple labels are joined with |, e.g. document_qa|answer_first. Missing labels become unknown; conflicting labels from multiple origins of identical content become mixed, never duplicated counts. Language is not inferred when absent from source metadata. Groups are mutually exclusive combinations, not arbitrary simultaneous marginal constraints.

Fixed-weight example:

{"method": "fixed", "group_fields": ["method"], "weights": {"original": 0.7, "translate": 0.2, "distill": 0.1}}

These are illustrative ratios. Weights must explicitly name every available training group; zero is allowed. Use a temperature plan first to inspect actual group names.

Temperature sampling uses p_i = n_i^alpha / sum(n_j^alpha): alpha=1 preserves scale, alpha=0.5 raises smaller groups’ relative share, and alpha=0 assigns uniform weights to nonempty groups. See XLM. The selected unit determines n; these probabilities are not quality scores.

CPT counts use the frozen target tokenizer. Token mixtures require identical tokenizer recipes across inputs. Samples are never truncated to fill quotas. Availability, whole-sample granularity, family caps and diversity filters can produce shortfalls. plan reports available/target/selected quantities, actual shares and shortfall; partial plans return exit code 1. build rejects shortfalls unless explicitly allowed, then preserves partial status. Selection is always without replacement.

Export contract

manifest.json: input identities, recipe, target/actual proportions, tokenizer and status.
records.jsonl: selection ID, content hash, group, original payload, known origins and document family.
CPT: train.jsonl, validation.jsonl.
SFT: also train.vision.jsonl, validation.vision.jsonl, original images/ and images.json.
checksums.sha256: complete file inventory and hashes.

Mixtures use finflow-mixture-dataset-v1, not the original SFT verifier. Load actual PIL images using training.mixture.load_vision_records(path, split="train"). Origins reference original sample/artifact IDs; complete historical auditing still needs the source data root. Snapshot admission is historical: this command does not query live policy revocations. Republish an approved snapshot when source policy changes.

Quality and diversity

Optional selection configuration:

{"selection": {"scores_file": "data/scores/quality.jsonl", "metric": "jaccard", "threshold": 0.9}}

Each score JSONL row requires mixture-pool id, content_hash, quality and complexity (finite 0–5). Cosine mode also requires finite, nonzero embedding vectors of identical dimension. Scores must cover the entire training pool, not just a quota-selected subset. Inspect IDs through training.mixture_data.load_pool(inputs, group_fields) or a sufficiently large unfiltered mixture’s records.jsonl.

Within each group, rank by quality × complexity, then reject candidates whose similarity to already selected examples meets the threshold. Family limits still apply. Jaccard uses full-text character trigrams, comparing assistant answers for SFT so repeated context does not dominate similarity. Cosine consumes externally computed scores/embeddings; no DEITA models are downloaded. This is DEITA-inspired; lexical similarity is not semantic similarity. Direct pairwise comparisons target bounded local pools; partition large datasets first.

Learned weights

Run DoReMi / RegMix, then copy config/mixture-learned.json and set the corresponding method and weights file with identical inputs/grouping. Learned CPT weights are applied as token quotas and bound to the exact frozen pool. Proxy training samples with replacement; static dataset export does not, so learned weights may still exceed available capacity.