Replace stale llama/mistral/phi model refs with models active on the
cluster: deepseek-r1 (1.5b, 7b-4bit, 0528-qwen3-8b-gguf), granite-4.1-8b,
qwen2.5 (3b, 7b), capybarahermes-2.5-mistral-7b, darwin-9b-opus. Update
benchmark_plans.py doc examples to match.
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.
- BenchmarkView.vue: convert from monolithic view to tabbed shell; each tab is
now its own component (ClassifierTab, CompareTab, LlmEvalTab, StyleTab, VoiceTab)
- StyleTab + VoiceTab: new benchmark modes for style and voice model evaluation
- app/style.py: FastAPI router for style imitation benchmarks
- app/voice.py: FastAPI router for voice benchmark endpoints
- scripts/benchmark_style.py + benchmark_voice.py: headless runner scripts
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.
Two bugs fixed:
1. Blank white page after vue SPA rebuild: browsers cached old index.html
referencing old asset hashes. Assets are deleted on rebuild, causing
404s for JS/CSS -> blank page. Fix: serve index.html with
Cache-Control: no-cache so browsers always fetch fresh HTML.
Hashed assets (/assets/chunk-abc123.js) remain cacheable forever.
2. Queue draining to empty on skip/discard: handleSkip and handleDiscard
never refilled the local queue buffer. After enough skips, store.current
went null and the empty state showed (blank-looking). Fix: both handlers
now call fetchBatch() when queue drops below 3, matching handleLabel.
Also: sync classifier_adapters LABELS to match current 10-label schema
(new_lead + hired, remove unrelated).
48 Python tests pass, 48 frontend tests pass.
