Brazenle/dom-analyst-compact-v1

Dom analyst compact v1 — experimental

This is Richard Jones's CPU-trained analysis-plan experiment, packaged for the optional expert workspace in his portfolio. It is a 494,032,768-parameter LoRA fine-tune of Qwen2.5-0.5B-Instruct, with q4 decoder weights and a shared INT8 vocabulary table. The ONNX graph is 361,958,538 bytes; the complete runtime package is approximately 378 MB. Download size does not establish peak browser memory use.

Semantic qualification failed. Review every proposed plan. Native PyTorch scored 7/24 exact complete plans; the preregistered exported q4 smoke scored 2/6. This is not a dependable autonomous analyst, and it does not write arbitrary R or React code. It proposes one of six structured analysis recipes with column indices. Validated R procedures perform the calculations; React renders the results.

The portfolio's default Dom chat uses a separate 16 KB command classifier, tested procedures, and project tools. This larger model is used only in the optional experimental workbench.

Provenance and changes

Base: Qwen/Qwen2.5-0.5B-Instruct at 7ae557604adf67be50417f59c2c2f167def9a775, Apache 2.0. CPU-only LoRA training used rank 16, alpha 32, 120 optimizer steps and 240 distinct synthetic examples, with no private dataset rows or external model teacher. Base weights were merged with the trained adapter and converted to ONNX q4. A later fixed-policy storage conversion replaced only the tied FP32 vocabulary table with per-row symmetric INT8 storage; it did not train a new model or improve semantic qualification. Other 457 initializers remained byte-identical. See NOTICE and evidence.json.

The compact graph preserved the original q4 generated tokens in all six compared cases, including incorrect answers. Two separate browser compatibility observations generated one expected plan and rejected a known incorrect plan before R execution. Those observations are not a fresh accuracy benchmark. Original FP32 export strict numerical parity missed its requested tolerance.

Runtime contract

Use Transformers.js 4.2 with CPU WebAssembly, dtype: 'q4' and session_options: {extra: {session: {disable_quant_qdq: '1'}}}. This optimizer compatibility option is required by the fixed shared-table experiment. Load the immutable repository revision recorded by the portfolio, never an unpinned branch. Retain the bundled training chat template with CRLF normalized to LF. Inputs contain a question and indexed column metadata, not data rows; the six-recipe protocol has a 384-token input bound and 96 generated-token limit. The portfolio validates output fields and column types before any execution.

release.json is the original frozen runtime release, including evidence recorded at packaging time. The later browser observations are described separately in evidence.json; neither changes qualified: false. There is no v2 checkpoint in this repository.

Limits

Evaluation used small synthetic schemas and agent-authored cases, not representative real users or production datasets. Wrong column selection and wrong task selection remain known failures. The model cannot establish causality, statistical validity, or whether a finding matters. A successful browser load only establishes compatibility. No claim is made of reduced RAM use, improved accuracy, faster responses, general R programming, or autonomous reliability.

manifest.json lists exact public payload sizes and SHA-256 hashes. Model weights contain no hosted inference service; inference runs in the visitor's browser after download. See the portfolio source for the validated tool boundary and review interface.