Adds benchmark_plans.py script, plans_bench API router, PlansBenchTab Vue
component, and registers /api/plans-bench in api.py. Also extends models
registry (cf-text catalog integration), cforch client, LlmEvalTab, and
ModelsView with cf-orch fleet support. Wires Planning mode into BenchmarkView.
- GET /api/train/jobs now returns {"jobs":[...]} instead of bare array
- GET /api/train/results now returns {"results":[...]} instead of bare array
- POST /api/train/jobs body key renamed config -> config_json to match Pydantic model
- SSE log handler now handles 'progress' event type (backend never emits 'log')
- Dashboard _get_active_jobs() adds model_key to SELECT and return dict
- corrections.py docstring updated: both /api/corrections and /api/sft prefixes noted
- test_train.py assertions updated to unwrap new envelope shapes
Replace 149-line api.py (with inline helpers, JSONL utilities, and ad-hoc
router registrations) with a 57-line pure factory. All business logic was
already extracted to domain modules in B1-B7; this removes the dead code
and adds the /api/corrections/* prefix alongside the /api/sft/* backward-
compat alias. Smoke tests updated to cover the new /api/corrections/ingest
and /api/dashboard routes.
Adds IngestRequest model and POST /api/sft/ingest route to
app/data/corrections.py. Sibling CF products (Peregrine, Kiwi, etc.)
can push pre-approved corrections via Bearer token auth
(AVOCET_INGESTION_SECRET). Records land as status=approved in both
sft_candidates.jsonl and sft_approved.jsonl immediately.
7 tests in tests/test_data_corrections.py cover 503 (secret unset),
401 (missing/malformed header), 403 (wrong secret), happy-path writes
to both files, and optional label field.
Replaces the ad-hoc _running_procs dict in api.py with a persistent,
inspectable SQLite job queue. Removes old /api/finetune/* routes and
_best_cuda_device from api.py. Adds /api/train/* routes (list, create,
get, cancel, run SSE, results). 16 new tests all passing.
Backend (app/imitate.py):
- GET /api/imitate/products — reads imitate: config, checks online status
- GET /api/imitate/products/{id}/sample — fetches real item from product API
- GET /api/imitate/run (SSE) — streams ollama responses for selected models
- POST /api/imitate/push-corrections — queues results in SFT corrections JSONL
Frontend (ImitateView.vue):
- Step 1: product picker grid (online/offline status, icon from config)
- Step 2: raw sample preview + editable prompt textarea
- Step 3: ollama model multi-select, temperature slider, SSE run with live log
- Step 4: response cards side by side, push to Corrections button
Wiring:
- app/api.py: include imitate_router at /api/imitate
- web/src/router: /imitate route + lazy import
- AppSidebar: Imitate nav entry (mirror icon)
- config/label_tool.yaml.example: imitate: section with peregrine example
- 16 unit tests (100% passing)
Also: BenchmarkView.vue Compare panel — side-by-side run diff for bench results
TaskEntry now includes prompt/system fields (default ""). Switch from
exact dict comparison to field-by-field assertions so the test is
forward-compatible with optional schema additions.
- sft.py: _DEFAULT_BENCH_RESULTS_DIR set to circuitforge-orch bench
results path; set_default_bench_results_dir() seam for test isolation
- test fixture resets default to tmp_path to avoid real-fs interference
- 136 tests passing
Closes#14
- _load_cforch_config() falls back to CF_ORCH_URL / CF_LICENSE_KEY /
OLLAMA_HOST / OLLAMA_MODEL env vars when label_tool.yaml cforch: key
is absent or empty (yaml wins when both present)
- CF_LICENSE_KEY forwarded to benchmark subprocess env so cf-orch agent
can authenticate without it appearing in command args
- GET /api/cforch/config endpoint — returns resolved connection state;
redacts license key (returns license_key_set bool only)
- SettingsView: connection status pill (cf-orch / Ollama / unconfigured)
loaded from /api/cforch/config on mount; shows env vs yaml source
- .env.example documenting all relevant vars
- config/label_tool.yaml.example: full cforch: section with all keys
- environment.yml: add circuitforge-core>=0.9.0 dependency
- .gitignore: add .env
- 4 new tests (17 total in test_cforch.py); 136 passing overall
Closes#10
- GET /api/models/lookup now returns compatible: bool and warning: str|null
- compatible=false + warning when pipeline_tag is absent (no task tag on HF)
or present but not in the supported adapter map
- Warning message names the unsupported pipeline_tag and lists supported types
- ModelsView: yellow compat-warning banner below preview description;
Add button relabels to "Add anyway" with muted styling when incompatible
- test_models: accept 405 for path-traversal DELETE tests (StaticFiles mount
returns 405 for non-GET methods when web/dist exists)
Adds optional failure_category to SubmitRequest and candidate records so
reviewers can classify why a model response was wrong, not just what to do
with it. Enables the fine-tune harness to filter training data by failure
type (e.g. exclude scoring artifacts, train only on genuine wrong answers).
Taxonomy: scoring_artifact | style_violation | partial_answer |
wrong_answer | format_error | hallucination
- app/sft.py: FailureCategory Literal type; SubmitRequest.failure_category;
stored on candidate record in POST /submit correct branch
- tests/test_sft.py: 3 new tests (stores value, null round-trip, 422 on invalid)
- stores/sft.ts: SftFailureCategory type exported; SftQueueItem + SftLastAction
updated; setLastAction accepts optional category param
- SftCard.vue: chip-group selector shown during correct/discard/flag flow;
two-step confirm for discard/flag reveals chips before emitting; category
forwarded in all emit payloads
- CorrectionsView.vue: handleCorrect/Discard/Flag accept and forward category
to POST /api/sft/submit body and store.setLastAction
- SftCard.test.ts: 11 new tests covering chip visibility, selection,
single-active enforcement, pending-action flow, emit payloads, cancel
Rename _strip_html/_extract_body to strip_html/extract_body (public API).
Remove duplicate _TextExtractor, strip_html, and _extract_body from
imap_fetch.py; import from app.utils instead. Update test_label_tool.py
to use the new public names.
- Delete app/label_tool.py (Streamlit UI retired; Vue SPA is sole UI)
- Extract _strip_html and _extract_body into app/utils.py (stdlib-only, reusable)
- Update tests/test_label_tool.py import to app.utils
- Rename start-api/stop-api/restart-api/open-api → start/stop/restart/open in manage.sh
- Remove STREAMLIT variable and all Streamlit-specific case blocks from manage.sh
- Update manage.sh usage section to reflect Vue+FastAPI-only commands
- Add data/sft_candidates.jsonl and data/sft_approved.jsonl to .gitignore
- Add sft.bench_results_dir key to config/label_tool.yaml.example
- except clause in cancel_benchmark/cancel_finetune narrowed from Exception
to _subprocess.TimeoutExpired (C1)
- _cancelled_jobs.discard() called after registering new proc to prevent
a stale flag from a prior run masking errors (I2)
- local `import subprocess` removed from run_benchmark and
run_finetune_endpoint; all Popen calls updated to _subprocess.Popen (I1)
- test patch targets updated from subprocess.Popen to app.api._subprocess.Popen;
cancelled-event tests updated to set flag in proc.wait() side-effect so
the discard-on-new-run logic is exercised correctly
Adds POST /api/benchmark/cancel and POST /api/finetune/cancel endpoints
that terminate the running subprocess (kill on 3s timeout), and updates
the run generators to emit a cancelled SSE event instead of error when
the job was intentionally stopped.
- load_and_prepare_data() now accepts Path | list[Path]; single-Path callers unchanged
- Dedup by MD5(subject + body[:100]); last file/row wins (lets later runs correct labels)
- Prints summary line when duplicates are dropped
- Added _EmailDataset (TorchDataset wrapper), run_finetune(), and argparse CLI
- run_finetune() saves model + tokenizer + training_info.json with score_files provenance
- Stratified split guard: val set size clamped to at least n_classes (handles tiny example data)
- 3 new unit tests (merge, last-write-wins dedup, single-Path compat) + 1 integration test
- All 16 tests pass (15 unit + 1 integration)
Implements load_and_prepare_data (JSONL ingestion with class filtering),
compute_class_weights (inverse-frequency, div-by-zero safe), compute_metrics_for_trainer
(macro F1 + accuracy), and WeightedTrainer.compute_loss (**kwargs-safe for
Transformers 4.38+ num_items_in_batch). All 12 tests pass.
