arcadia-impact/scimt-dispatch-aft-v1

Dispatch true-midtraining AFT data and run evidence

This repository is the public data and provenance companion to jbostock/scimt-dispatch-models-v1. It contains synthetic Dispatch episodes, exact launch/training manifests, raw model generations, deterministic scores, logs, and publication receipts for a study of whether different midtraining histories select different policies after byte-identical, objective-ambiguous supervised fine-tuning.

It is not just a conventional row-only Hugging Face dataset. It is organized as timestamped, self-contained experiment records so a result can be traced back to its data, code state, parent checkpoints, configuration, and raw outputs.

The experiment in brief

Dispatch is a synthetic logistics decision task in an invented world. The Coin policy maximizes a plan's coin total, while the Charter policy applies a fixed compositional rulebook. The two policies agree on every AFT training demonstration, and neither objective is named in the prompt. They disagree on the held-out conflict set, which is used to measure which rule each model generalizes.

Two Gemma 3 12B parents received different Coin or Charter midtraining histories and then the same 100M-token general SFT stage. The parents subsequently receive the same ordered 2,048 agreement-only AFT rows.

Repository layout

runs/<RUN_ID>/
├── launch/                 # source manifest and launch configuration
├── data/episodes/
│   ├── episode_manifest.json
│   ├── episodes/           # canonical train/eval episode records
│   └── datasets/           # trainer-ready JSONL variants
├── evaluation/
│   ├── samples/            # raw generations for every endpoint/split
│   └── ...                 # detailed and aggregate deterministic scores
└── evidence/               # resolved provenance, package/GPU state, logs,
                            # summaries, publication receipts, terminal marker

The model repository separately stores the LoRA adapters and training-side configs/traces at runs/<RUN_ID>/<coin|charter>/....

Main row files

path below runs/<RUN_ID>/data/episodes/	rows	purpose
episodes/train_agreement.jsonl	2,048	canonical agreement-only AFT training episodes
episodes/eval_agreement.jsonl	512	disjoint held-out agreement evaluation
episodes/eval_conflict.jsonl	512	disjoint held-out policy-conflict evaluation
datasets/aft_agreement.jsonl	2,048	trainer-ready agreement-only conversations
datasets/aft_conflict_balanced.jsonl	2,048	balanced conflict control, retained but not used for this AFT gate
datasets/aft_mixed_coin.jsonl	2,048	Coin-labeled mixture control, retained but not used for this gate
datasets/aft_mixed_charter.jsonl	2,048	Charter-labeled mixture control, retained but not used for this gate

The experiment trains only on aft_agreement.jsonl. The other generated variants are retained because the audited generator emits a complete family of controls; their presence does not mean they were included in the training mix.

Runs

run	training horizon	state	notes
20260807T100738Z	64 steps / 1 epoch	complete	SFT baseline plus checkpoints 4–64 evaluated
20260807T104104Z	128 steps / 2 epochs	complete	SFT baseline plus checkpoints 4–128 evaluated
20260807T110710Z	2,048 steps / 32 epochs	complete	checkpoints 4–2,048 plus generic controls; retained in the consolidated model repository

Earlier prefixes such as 20260807T093541Z, 20260807T094026Z, 20260807T094440Z, and 20260807T094843Z are preserved aborted/failed launch records. They are provenance for integration failures, not completed scientific runs; consult their ABORTED.json or FAILED.json markers.

The completed 64- and 128-step runs intentionally contain their own copies of the generated data so each timestamped record is independently auditable. The agreement training JSONL has SHA-256 2220d77d4e6256aec4b67f096576d56d779336a14ddea420a0c8734b6afa616b in both runs. Train/eval prompt overlap and scenario overlap are zero.

Data generation and fields

The generator uses seed 314159 and the established four-crew, one-run Dispatch SDF-v1 design. Canonical episode records include the rendered prompt, scenario identifiers and factors, candidate plans, the Coin and Charter target choices, whether the policies agree or conflict, and audit metadata. Trainer- ready rows render the selected target as a chat conversation.

Prompts are synthetic and use an invented setting, names, rules, and logistics facts. No personal data is intentionally included.

Evaluation records

For every parent/checkpoint endpoint, the repository retains raw greedy generations on 512 agreement and 512 conflict episodes. Scoring is deterministic and reports:

• held-out agreement accuracy;
• Charter, Coin, and Other/malformed rates on conflict episodes;
• episode-level Wilson intervals and detailed scorer output;
• model/tokenizer view and tokenization checks.

The completed generic control evaluates the unchanged SFT parents and every retained adapter on a fixed 40-question MMLU plus 40-question GSM8K subset. It also records judge-free diagnostics for empty responses, parseability, truncation, repeated four-grams, exact-response duplication, and accidental Dispatch-language intrusion. Results live under the corresponding training run's generic_eval/20260807T135326Z prefix.

Reproducibility contract

Each completed run records, at minimum:

• exact Git commit and tree plus a per-file source-manifest digest;
• immutable parent repository revision and verified file identities/sizes;
• dataset hashes and generator audit;
• fully resolved Axolotl configurations and training contract;
• data/training/evaluation seed (314159);
• package locks, GPU inventory, complete training traces, and stdout/stderr;
• raw evaluation samples and aggregate summaries;
• immutable model/log publication revisions and a terminal completion marker.

Use the revision recorded in evidence/publication.json or evidence/RUN_COMPLETE.json when citing a run, rather than relying on moving main.

Limitations and responsible use

• This is a synthetic alignment research artifact, not a representative corpus of real logistics decisions or human preferences.
• Completed results currently use a single AFT seed. Episode-level confidence intervals do not capture training-run variance.
• Coin and Charter histories differ in content and complexity, so the two-arm comparison does not identify complexity alone.
• The presence of trainer-ready conflict/mixed controls must not be confused with the agreement-only data actually used by this gate.
• Long-horizon AFT may overfit or collapse even when training loss is finite; interpret results together with conflict and generic-capability evaluations.

Related artifacts

• Consolidated midtraining, SFT, and long-run AFT checkpoints: jbostock/scimt-dispatch-models-v1
• Experiment specification and code: science-of-midtraining PR #420
• Closest conceptual predecessor: Li et al., Model Spec Midtraining (2026)