G33-k/minicpm5-brand-tools-eval-kit

MiniCPM5 brand-tools training and evaluation kit

Version 2.0 · 12 September 2026 · synthetic, offline evidence worlds

This kit is for the two existing functions verify_company_website and find_customer_facing_pages. Their TypeScript implementations and function schemas are copied unchanged from the prior brand-tools package. The kit creates new synthetic environments and separates training targets from model-visible evaluation prompts and private grader data.

Dataset

Partition	Episodes	Assistant decisions	Intended use
Train	1,032	2,856	Supervised fine-tuning
Validation	258	714	Validation loss, checkpoint selection, development evaluations
Test	386	1,130	Final evaluation after model/prompt settings are frozen
Total	1,676	4,700

The test partition contains 258 in-distribution cases and 128 challenge cases. There are 37 recurring scenario families and 16 test-only challenge compositions. Company identities, primary domains, and user-prompt template IDs are disjoint across partitions. Shared schemas, policies, task primitives and page-content templates are intentional. This is not a human-curated real-world benchmark and not proof of model accuracy.

Start with the offline checks

Requires Node.js >=22.16 and Python >=3.10. The dataset generator, tool replays, and graders require no npm packages, credentials, network, or GPU.

npm test
python scripts/audit.py
python -m unittest discover -s tests -p 'test_*.py'
(cd training && python -m unittest test_prepare.py)

npm test replays all 1,676 transcripts, runs negative scoring tests, and runs a local HTTP contract mock. The HTTP mock is not a model evaluation.

What to load

File	What it contains	Model exposure
sft/train.jsonl	Training conversations with correct calls and final responses	Training inputs and targets
sft/validation.jsonl	Separate validation conversations	Validation only, never training loss
eval/validation.prompts.jsonl	Initial prompts for development rollouts	Yes
eval/test.prompts.jsonl	Initial prompts for final rollouts	Yes, at final evaluation
eval/*.decisions.jsonl	Teacher-forced conversation prefixes for next-action tests	Yes
grader/*.decisions.gold.jsonl	Expected next actions	No
grader/*.cases.jsonl	Environment setup and private scenario metadata	No
grader/*.worlds.jsonl.gz	Synthetic source snapshots and page-intent labels	No
grader/test.references.jsonl	Reference test trajectories, for grader audit	No; do not train on this

Do not glob all JSONL files into a training loader. Use only the two explicitly named SFT files. training/prepare.py deliberately reads train and validation, not test. Validation targets are used to calculate validation loss, never model weight updates in the supplied trainer.

Fine-tuning preparation

The existing native-template preparation, LoRA training and merge helpers are included. The preparation helper has been restricted to the two SFT partitions. Install the GPU-compatible PyTorch build separately, then the training requirements.

python -m pip install -r training/requirements.txt
python training/prepare.py --data sft --out data/tokenized --max-length 16384
python training/train_lora.py --data data/tokenized --out runs/minicpm5-brand-tools

Preparation resolves and records the checkpoint revision and tokenizer-template hash; it preserves tool calls, masks non-assistant targets, and fails on incompatible rendering or overlength examples. The native tokenizer, GPU training, and model serving have not been executed for this delivery. The

Status update (2026-09-21, this checkout). The statements above describe the originally shipped snapshot. The corpus has since been regenerated twice — the 2026-09-18 gate rework and the 2026-09-21 disambiguation-policy change — and this checkout carries the latter, so the counts above are current but the "no training performed" statements are not. Since then: the native MiniCPM tokenizer has been run (longest decision 9,809 tokens, making 16,384 a verified bound for this revision), MiniCPM5-2B was trained on GPU and served through a local OpenAI-compatible endpoint, and a full validation evaluation was run against the previous revision (252/258 episodes, training-runs/reports/unsloth-eval-2026-09-20.md, revision-pinned and retired by the policy change). Current run state: training-runs/HANDOFF.md. Evaluation contract: training-runs/EVAL-PLAN.md. 16,384-token argument is a chosen limit, not a verified bound on this corpus. Read docs/EVALUATION_GUIDE.md and docs/DATA_CARD.md before training.

Evaluate an actual served model

The endpoint must already produce standard OpenAI-compatible tool_calls for this checkpoint. Plain-text native function tags are not automatically parsed by this runner. Supply the serving layer's model identifier; no provider is assumed. The runner sends only messages and public tool schemas, not grader files.

node --experimental-strip-types scripts/evaluate.ts \
  --mode model --split validation \
  --base-url http://127.0.0.1:30000/v1 --model YOUR_SERVED_MODEL_ID \
  --out reports/validation-model.json

After selecting the checkpoint and freezing the prompt/configuration:

node --experimental-strip-types scripts/evaluate.ts \
  --mode model --split test \
  --base-url http://127.0.0.1:30000/v1 --model YOUR_SERVED_MODEL_ID \
  --out reports/test-model.json

Optional authorization is read from MODEL_API_KEY by the runner; never place secrets inside datasets. The runner supports --temperature, --max-tokens, --max-steps, --timeout-ms, --limit, and --template-kwargs off for endpoints that do not accept template controls. Do not use --limit results as full-test scores. By default, decoding temperature is zero and thinking is disabled via chat_template_kwargs; actual server behavior must be verified by the operator.

To sanity-check the harness without a model:

node --experimental-strip-types scripts/evaluate.ts \
  --mode reference --split test --out reports/test-reference.json

That mode uses a scripted reference policy. Its report states is_model_result: false. A 100% reference result means fixtures and grading rules are consistent; it says nothing about MiniCPM accuracy.

Score standalone next-action predictions

Prediction format is one JSONL record per decision:

{"id":"COPY_THE_DECISION_ID","message":{"role":"assistant","content":"...","tool_calls":[]}}

For function calls, include a standard tool_calls entry. arguments can be a JSON string or object. Run:

python scripts/score_decisions.py --split validation \
  --predictions your-validation-predictions.jsonl \
  --out reports/validation-decisions.json

Missing predictions count as failures; duplicate and unknown IDs are reported. This measures next-action correctness given reference history, not autonomous end-to-end success.

Reproduce the datasets

npm run build:data
python scripts/audit.py
npm test

Generation is deterministic. Runtime-issued verification IDs are randomized per model rollout so a stored token cannot be memorized and replayed as authorization. Source snapshot hashes and checksums.sha256 provide artifact integrity checks.

Limits

All 1,676 companies and their registration/domain evidence are fictional. Zero live-company cases, zero human-reviewed labels, zero actual model rollouts, and zero fine-tuning runs are claimed. The unchanged tools are precision-oriented static-HTML heuristics; this benchmark does not establish global company-registry coverage, fraud detection, browser-rendering quality, or real-world source recall. The benchmark has two tools only; it does not test brand profiling or copywriting.

Validation evidence (corpus revision 2026-09-21)

validation-2026-09-21/ holds the full 258-episode adapter-vs-base sweep: per-arm shard reports, the merged 258-episode reports, the harness comparison table, the run manifest, the two family-aligned shard case files, the 8 arm rollout traces, and validation-progress.log. Both adapters scored 258/258 episodes (scenario macro 1.000) against untouched bases at 0/258, served as the unmerged LoRA over the same NF4 base, greedy with thinking off. CHECKSUMS.sha256 covers every file in that folder. The adapters themselves live at G33-k/minicpm5-2b-brand-tools-controller-lora and G33-k/qwen3-vl-4b-brand-tools-controller-lora.