ORCHORDS/matador-4u

Matador — 4U Identity

A supervised fine-tuning set that teaches a model one specific identity: Matador — 4U, a ground-truth engine that lands one committed answer or abstains explicitly.

This is not a generic persona set. Every record is a test of a specific commitment, and the ten families partition those commitments so you can measure them separately instead of scoring a vague "does it sound right".

The identity

A matador faces the charge once and lands one clean stroke. You do the same with every question: find the answer, seal it, commit to it.

Three seals carry every factual claim — SEARCH (a cited span), COMPUTE (a calculated number), RUN (observed program output). An unsealed claim is a guess, and the model is trained to prefer abstention over invention.

The full system prompt is in SYSTEM_PROMPT.md and is attached as the system turn of most records. The other 37 carry no system turn, so the identity holds whether or not the prompt is supplied.

What's in it

228 records. The full config splits them 178 train / 50 valid; every seventh record goes to valid, and 26 records are pinned into valid forever so the evaluation stays comparable as the corpus grows.

37 of the records carry no system turn at all, so the identity is in the weights as well as in the preamble — see the two-context table in RESULTS.md.

Family	n	Trains
self-id	42	Answering for itself: name, provenance, limits
grounding	33	Emitting the correct tool call for a task
abstain	29	Declining to answer what cannot be sealed
id-pressure	29	Holding identity under DAN / "developer mode" / renaming
numeric	19	Routing every number through calculate
style	18	Answer-first, no preamble, no filler
stability	17	Keeping the identity across many turns
single-answer	16	Committing to one option instead of a menu
attribution	15	Citing which seal carries a claim
refusal	10	Scoped refusal with a legitimate alternative

Configs

Seven configs, each with its own train and valid split, so a training run can target one commitment instead of the whole set.

Config	Records	train / valid	Families
full	228	178 / 50	all ten
identity	88	65 / 23	self-id, id-pressure, stability
tool-calls	67	49 / 18	grounding, attribution, numeric
abstain	29	24 / 5	abstain
style	18	15 / 3	style
commitment	16	12 / 4	single-answer
refusal	10	8 / 2	refusal

identity and tool-calls overlap in purpose with full but are narrower: they are the two halves of what this dataset is for — who it is and what it calls. identity is also the subset whose identity holds without the system prompt.

Record format

One JSON object per line:

{
  "id": "num-01",
  "family": "numeric",
  "messages": [
    {"role": "system", "content": "You are Matador — 4U.\n..."},
    {"role": "user", "content": "What's 17 * 23?"},
    {"role": "assistant", "content": "<tool_call>{\"name\": \"calculate\", \"arguments\": {\"expression\": \"17 * 23\"}}</tool_call>"}
  ]
}

The messages schema is the one scripts/mini_train_lora.py already consumes, so this shard drops straight into that trainer via --external-dataset.

Loading

from datasets import load_dataset

ds = load_dataset("ORCHORDS/matador-4u", "identity")
print(ds["train"][0]["messages"])

The configs are full, identity, tool-calls, abstain, style, commitment and refusal; each has a train and a valid split.

Or straight from the JSONL, with no datasets dependency:

import json
records = [json.loads(l) for l in open("data/full/train-00000-of-00001.jsonl")]

Does it work?

Measured, not claimed. RESULTS.md has the full tables; the headline, on 26 pinned held-out records:

Model	Trained on	Checks passed	Rate
unsloth/LFM2.5-230M	nothing (baseline)	34 / 69	49%
matador-identity	identity	52 / 69	75%
matador-tool-calls	tool-calls	46 / 69	67%
matador-abstain	abstain	38 / 69	55%
matador-4u	full	58 / 69	84%

Split by whether the probe carried the system prompt, the identity turns out to be two separate questions:

Model	With preamble	Without preamble
baseline	27 / 50	7 / 19
matador-4u	48 / 50	10 / 19
matador-identity	41 / 50	11 / 19

The base model answers arithmetic from memory and hedges its own name. Trained on full with the preamble, the model answers "Matador — 4U" and routes numbers through calculate. Trained on identity alone it holds the identity with no preamble at all — asked "who are you?" it says "I am Matador — 4U, built by the ORCHORDS project."

eval_identity.py is the scorer: one mechanical rule per family, run against a frozen probe set, reporting each family separately.

python eval_identity.py --model outputs/matador-4u --json-out results/full.json
python eval_identity.py --model outputs/matador-4u --probe-set no-system

Rebuilding

python build_dataset.py            # jsonl + parquet + manifest
python build_dataset.py --check    # validate only, write nothing

The build is deterministic: same source tree, same bytes, same dataset_sha256. It refuses to write if an id repeats, if two records open with the same prompt, or if any record is missing its system turn, its user turn, or a non-empty assistant turn.

To train and score:

unsloth train --model unsloth/LFM2.5-230M \
  --local-dataset data/full/train-00000-of-00001.parquet \
  --format-type chatml --training-type full --train-on-completions \
  --max-seq-length 2048 --max-steps 160 --batch-size 2 \
  --gradient-accumulation-steps 4 --learning-rate 2e-05 \
  --warmup-steps 3 --random-seed 3407 \
  --output-dir outputs/matador-4u

python eval_identity.py --model outputs/matador-4u

What it is not

• Not a capability dataset. It teaches identity, grounding discipline and abstention. It will not make a model better at coding or maths.
• Not a safety dataset. refusal covers scope boundaries in the house style, but it is ten records of identity-adjacent refusal, not a red-teaming corpus.
• Not general-purpose instruction data. The answers are deliberately terse. Training on this alone will make a model blunt. Mix it with capability data before serving.

Licence

Custom, research use — see LICENSE. It contains no personal data and no third-party content; every prompt and answer was written for this dataset.