Skip to main content
Industry recipes use the same SFT pipeline with separate data and adapters. They are task templates, not pretrained domain agents or proof of business outcomes. The current checkout adds deterministic synthetic corpora for ecommerce, logistics, and SaaS support.

​
Generate and audit without compute

Generation, audit, and these previews use no model or provider. Use a new output directory. Default per-domain counts are 512 train, 64 validation, and 128 reserved test conversations across eight balanced families per split. Counts can be changed in multiples of eight; generating data does not train weights. Each domain retains train/validation/test JSONL, private synthetic cases, and a training manifest; the root manifest binds all artifacts. The audit reconstructs the corpora, replays teacher actions, and checks provenance/hashes. Declared status is synthetic: true and not_independently_reviewed. No actual shipment, refund, or account setting changes.

​
Preserve split roles

Use only train.jsonl for optimization. Validation supports selection; reserve test for a separately frozen assessment. The studio, packaged SFT loader, industry training, and local-source provider preflights reject declared held-out or ambiguous training rows. Top-level split and metadata.split must agree; duplicate JSON keys fail the read. Renaming a held-out file does not change its declared role. Removing or falsifying annotations cannot be detected from an isolated file, so retain and audit the complete bundle. Exact source/case overlap checks do not prove semantic or temporal independence. The current teacher-forced industry evaluator continues to reject synthetic data for its quality gate.

​
Prepare a reviewed source project

The initialized examples are synthetic demonstrations. Replace or expand them with reviewed source conversations before making quality claims. Preparation writes new split files and a manifest; editing the files afterwards invalidates training. Use a new project directory after edits. Supply --model-revision with the real immutable model commit for reproducible tokenizer and weight loading. Records require text messages ending in an assistant target. The industry contract also supports developer/tool roles and OpenAI-style tools/tool calls with matching result IDs; this is broader than the studio’s text-only alternating chat form. Multimodal message content is outside this workflow. The model’s actual chat template must support the data. Use a stable group_id for the same case/customer/source conversation. Exact duplicates are removed, and matching initial user requests or group IDs stay together, including transitive connections. At least two independent groups are required. The split fraction applies to groups, so row proportions can differ.

​
Execute and inspect lineage

On a configured CUDA host with the training dependencies, remove --dry-run to execute the industry SFT trainer. CPU previews remain planned; they never report a trained adapter. Dataset split seed and training seed have different roles; retain both. Local SFT captures source bytes, parsed training content, source-line counts, and effective rows before model loading. Industry training checks that capture against the project manifest, while studio workers check their reviewed plan. The export retains source/content lineage even if the original file is later changed or deleted. Source-free in-memory inputs remain explicitly unbound to a file. The run becomes trained only after the fixed-rank LoRA safetensors export, tokenizer metadata, and lineage pass structural checks. Missing shards, incomplete tensor pairs, corrupt/empty headers, or mismatched lineage fail completion. Structural validation checks metadata and tensor layout, not numerical finiteness, optimizer-update counts, or inference quality. quality_gain_verified remains false. Evaluate against reviewed independent data before promotion. Current operational facts belong in tools/retrieval; examples teach how to use them. Keep synthetic corpus checks separate from the retained River learning studies. Source basis: the October 20, 2026 docs/INDUSTRY_FINETUNING.md and packaged SFT/dataset contracts. Continue with the studio or paired checkpoint evaluation.
Last modified on October 20, 2026