Backend (app/api.py):
- GET /api/benchmark/models — returns installed models grouped by adapter
type (ZeroShotAdapter, RerankerAdapter, GenerationAdapter, Unknown);
reads _MODELS_DIR via app.models so test overrides are respected
- GET /api/benchmark/run — add model_names query param (comma-separated);
when set, passes --models <names...> to benchmark_classifier.py
- GET /api/stats — add benchmark_results field from benchmark_results.json
Frontend:
- BenchmarkView: collapsible Model Selection panel with per-category
checkboxes, select-all per category (supports indeterminate state),
collapsed summary badge ("All models (N)" or "N of M selected");
model_names only sent when a strict subset is selected
- StatsView: Benchmark Results table (accuracy, macro_f1, weighted_f1)
with best-model highlighting per metric; hidden when no results exist
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
- manage.sh: dev command starts uvicorn --reload on :8503 and Vite dev
server (auto-port from 5173); kills API on EXIT/INT/TERM trap
- manage.sh: ENV_UI defaults to 'cf' env (overridable via AVOCET_ENV)
- vite.config.ts: add server.proxy to forward /api to :8503 so Vite
dev server can reach the backend without CORS issues
- sft.py GET /config: use `or {}` guard so `sft: ~` (null YAML) doesn't
return None instead of the default empty config
- CorrectionsView: convert handleCorrect/Discard/Flag and handleUndo from
optimistic to pessimistic — queue mutation only happens after server
confirms; failures leave item in queue so user can retry cleanly
- SettingsView: call loadSftConfig() on mount so saved bench_results_dir
is populated instead of always starting empty
- Add /corrections route to Vue router (lazy-loaded CorrectionsView)
- Add Corrections nav item (✍️) to AppSidebar after Benchmark
- Add cf-orch Integration section to SettingsView with bench_results_dir
field, run scanner, and per-run import table
- Add GET /api/sft/config and POST /api/sft/config endpoints to app/sft.py
- Add v-if guard on failure-reason <p> so null renders no element (not literal "null")
- Clarify mid-quality test description: score is 0.4 to <0.7 (exclusive upper bound)
- Add test: renders nothing for failure_reason when null (+1 → 14 SftCard tests)
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.
Reentrant gradient checkpointing (the default) conflicts with Accelerate's
gradient accumulation context manager -- causes 'backward through graph a
second time' on the first training step. use_reentrant=False uses the
non-reentrant autograd hook path which is compatible with Accelerate >= 0.27.
- 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.
