zack-r-s/hackring-pyat-control

Hackring PyAT control data

Versioned synthetic data for a 24-PV FODO storage-ring simulation. The v001 source contains Xopt/PyAT read-write-read episodes and GPT-authored operator descriptions. The v002 source contains 512 deterministic PyAT teacher actions, 512 checked GPT prompt paraphrases, 640 canonical SFT conversations, and 1,152 final SFT conversations (864 train, 144 validation, 144 test). That final v002 split is the input to local training pass training/v001, published as model v002. Version v003 combines 640 canonical Osprey conversations with 480 new local-PyAT conversations, for 864 train, 128 validation, and 128 test. Its new measured actions use seeded error kicks and share the same fixed 24-PV ring; no new GPT pass was made. The v003 split is the input to local training pass training/v002, published as model v003.

v003/*.jsonl, v002/sft/*.jsonl and v002/sft-canonical/*.jsonl contain messages arrays with system, user, assistant, and tool turns. The assistant emits the base model's tool-call syntax; tool responses and measurements come from the simulator. v002/raw/actions.jsonl contains measured source actions, and v002/annotations/gpt-v001/paraphrases.jsonl contains language variants. Manifests and indexes record hashes and row lineage. v001/annotations/direct-v003/ is the selected completed v001 annotation pass; earlier pilots are omitted.

v004 skill-routing pass

The v004 split has 432 train, 72 validation, and 72 test conversations from fresh seed-disjoint local PyAT disturbances. Unlike v002/v003, the model target is orbit_correct(plane, effort), not numerical corrector setpoints. A deterministic bounded response-matrix solver computes the writes, and the conversation includes measured PyAT after-state. Training prompts cover gentle, balanced, and aggressive intent, including requests with no effort preference that mean balanced. No GPT pass was used. See v004/README.md, v004/manifest.json, and the matching model v004.

Mark III evaluation

PR #8 includes a completed Mark III validation and evaluation. The immutable v001–v004 manifests and conversation exports were revalidated without regenerating or relabeling data. The results index binds every local PyAT and Gym report to pinned model revisions, the checked-in 28 correction cases, direct/skill task hashes, and measured outcomes. All seven full Gym runs contain 28 tasks × four repeats with zero rollout failures. Direct numerical channel_write and orbit_correct(plane, effort) are different interfaces and have separate tables.

In local greedy MPS evaluation on 56 mixed read/write/correction prompts, base 230M passed 1, model v002 passed 24, model v003 passed 26, and base 1.2B Instruct passed 0. On the correction-only Gym direct task, v002 met the measured goal in 12/112 rollouts and v003 in 6/112, so the local ordering does not replicate. On the Gym skill task, model v004 met the measured goal in 72/112 rollouts versus 73/112 for base 230M; v004 selected the exact plane and effort in 42/112 versus 31/112. This fixed task set does not establish a v004 deployed-success advantage over base. The skill includes a deterministic bounded solver; success is not evidence that the model computed kicks itself.

Mark I remains the pre-reconciliation record, and Mark II was a partial diagnostic. Mark II's v004 Gym run used a direct-write task that cannot score its skill interface. Mark III replaces that comparison with separate frozen tasks and one-attempt server enforcement. The shared live fixture and browser UI passed smoke checks; exact numeric writes use a labeled bounded harness path.

This is a simulated plant, not real accelerator telemetry. GPT wrote prompt language only; measured PV values and actions originate from PyAT and the deterministic teacher. Splits should remain separate when evaluating. Exact values and channels are particular to this 24-PV ring; no real-machine safety or generalization is established.

Source code and training setup: https://github.com/zack-r-s/nvidia-redhat-small-models (integration PR #8). Versioned model checkpoints: https://huggingface.co/zack-r-s/hackring-LFM2.5-230M (dataset/model tags v002, v003, and v004). See the training and release chart.

The v003 tag freezes the mixed-backend training pass and its source lineage. The v003 local error-kick seeds are disjoint from the frozen v001 and v002 PyAT evaluations.

Use each matching dataset/model tag: v002, v003, or v004. These tags preserve each pass even as main advances.