munod/tachyone_slm_mixture_v1

Tachyone-SLM-Mixture-v1

Synthetic SFT training data for the TachyOne SLM — the local System-2 of the TachyOne System-1/System-2 hybrid, a Qwen2.5-0.5B-Instruct model fine-tuned with LoRA to answer typed decision questions (choice / score / noul) over the frozen /v1/systemone wire protocol.

• 50,180 clean records (45,175 train / 5,005 validation) and 50,180 ready-to-train SFT pairs
• 7 languages × 5 domains · labels derived from text (ADR-0014/ADR-0015)
• Contamination audit: PASS — overlap 0/0 against the 18,000 frozen eval records
• Generated by the existing pipeline of munod/tachyone (imported, never modified), config + seed pinned

Product specification: munod/tachyone_slm docs — PRD v1.2.0 §3.3/§3.4. Full dataset card with caveats: docs/slm-mixture-v1.md in that repo.

Usage

from datasets import load_dataset

# raw wire records: state + question (type/instructions/criteria) + gold target
records = load_dataset("munod/tachyone_slm_mixture_v1", "records")

# rendered SFT pairs: prompt (wire request + "Answer:") -> gold label
pairs = load_dataset("munod/tachyone_slm_mixture_v1", "pairs")

Every prompt is the canonical §5.1 request re-parsed by tachyone.wire.parse_request; every target is a member of its own candidate set:

type	target	scoring use (PRD §3.4)
choice	criteria key ("technical")	softmax over label logprobs
score	level index as string ("2")	softmax over levels, expected value Σ i·pᵢ
noul	"true" / "false"	sigmoid(ℓ(true) − ℓ(false))

Splits

Config	Split	Rows
records / pairs	train	45,175
records / pairs	validation	5,005

The split is a pure function of the record bytes — val iff int(sha256(raw_line)[:8], 16) % 10 == 0 — so membership is stable across runs and independent of record order, ids (positional, colliding across domains) and the language/domain cycles. The validation split is reserved for temperature fitting (never trained on, never an eval set).

Provenance & integrity

Field	Value
Config	training/configs/data_sft_slm.json (sha256 33802ba0…3fe9009d)
Seed	20261003 (outside {1 train, 2 eval, 42 default})
Generator	training/generate_data.py of munod/tachyone (unmodified)
Raw → clean	85,500 → 50,180 (41.31% removed by the contamination dedup)
Raw generation sha256	563d0c631fa35aed19e4a93ab1f3b4237e7544ef5646f347650fbdf06d27df15
Clean records sha256	9cf2e1463348d3333cce5484106e2bebff37303b3656ae67fde4e43d51369491
Train sha256	ac473873ee731b7cb55d3e0a7cf92b0a91b9372551a8c5263227b6c901ea8f78
Validation sha256	c8d6d3c61d17993edf84eabadbf959b294c7675b979df2bf25c17700fa6a6fa5
Pairs sha256	7924a0d54939657be7b778610f7ed33955bc750099cdcb9e62df632294480d82

The machine-readable evidence ships in provenance/: manifest.json (config + counts + per-cell table), split_render.json (dedup/split/render reports), audit.json (contamination audit) and data_sft_slm.json (the config itself).

Contamination handling (why 85,500 → 50,180)

The generator's phrase-bank pools are shared with the frozen eval sets, so a different seed alone does not isolate the datasets: 41.31% of raw records matched a frozen row. Those records were removed (state-text and prompt-fingerprint match against eval_en, eval_multi, eval_en_domains, eval_multi_domains — 18,000 records) and the audit re-run:

Audited side	state texts	prompt fingerprints
clean dataset (50,180)	0	0
temperature-fit val (5,005)	0	0

Public probes (typed-decisions, MASSIVE, XNLI) were absent locally and are recorded as skipped with the structural reason: the generator's only inputs are its phrase banks plus training/data/domains/*.json — it has no external dataset import path.

Coverage

All 35 language × domain cells are present (en, pt, es, fr, de, it, nl × support, ecommerce, agent_tools, documents, voice), 451–1,837 records each (see provenance/manifest.json). By primitive after dedup: choice 19,131 · score 17,442 · noul 13,607 — noul is the thinnest primitive, gate on it when training.

Caveats

• Synthetic and templated. Frozen eval content is excluded (overlap 0), but the template structure is shared by construction — the same generator produces both. Expect easier generalization within this templating than to real traffic; public probes remain the external check.
• raw dataset is not published — the pre-dedup file contains frozen eval rows by design and is reproducible locally from the config in provenance/.
• Not a benchmark: never train on the validation split or report accuracy from it.

License

Apache-2.0 — inherited from munod/tachyone (see LICENSE).