Repository navigation
Release Notes — v1.2.0
Release date: 2026-09-29
Distillery expands from a single causal-LM fine-tuner into a multi-kind task-to-model platform: 10 model kinds, a real llama.cpp serving backend, and schema-constrained decoding.
✨ Added
New model kinds (1 → 10)
| Kind | What it does | PR |
|---|---|---|
seq_classifier |
Text classification — label + confidence scores (sentiment, intent, topic, spam) | #3 |
token_classifier |
NER / token classification — labeled spans (people, orgs, dates, PII, custom) | #5 |
embedding / reranker |
Bi-encoder embeddings + cross-encoder rerankers; /embed, /rerank, /embed-corpus; nDCG@10 / MRR / recall base-vs-tuned |
#4 |
preference_lm |
Preference alignment via DPO from (chosen, rejected) pairs | #6 |
vision_lm |
Document AI with Qwen2-VL — ZIP/PDF import (auto-rasterized), multipart /predict, field-level exact-match/F1, ANLS, JSON validity |
#7 |
asr |
Whisper speech-to-text — speaker-grouped splits, WER/CER vs base, domain-glossary recall, /transcribe |
#8 |
tabular |
Gradient-boosted classification/regression on CSV (LightGBM / XGBoost / CatBoost, numpy fallback); SHAP, Platt calibration, leakage warnings | — |
time_series |
Forecasting with 80% intervals — seasonal naive / drift / SES / Holt-Winters / lag-GBM; rolling-origin backtests ranked by MASE; /forecast |
— |
Platform & infrastructure
- Common training infrastructure — Python trainer refactored into task modules (
trainer/tasks/{sft,classifier,ner,embedding,reranker,dpo,vision,asr,tabular,forecast}.py) behind a single config schema; per-kind dataset registry with validation, stats and bulk importers. - llama.cpp inference backend (
INFERENCE_BACKEND=llamacpp) — managedllama-serverprocess per deployment: health-checked, auto-restarted, stopped viaPOST /deployments/{id}/stop. - Constrained decoding — JSON Schema → GBNF grammar generation;
PredictConstrainedforces schema-valid JSON output for extraction tasks. - Local blob store (
internal/infra/blob) — stores images, rasterized PDF pages and audio;GET /api/v1/tasks/{taskID}/blobs/{key}. - PDF rasterizer — one rendered page image per example (
internal/infra/pdfraster,trainer/pdf_rasterize.py).
Web UI
- New
/tabular.htmlpage — CSV upload, column mapper with leakage/frequency hints, training form, live progress, leaderboard, importance, ROC / confusion matrix / residuals, backtest chart, batch predict/forecast. - Per-kind dataset forms and metric renderings (WER, nDCG, F1, MASE, …) across Tasks and Models pages.
📚 Documentation
- New guides: embeddings-rerankers, vision-language-document-ai, speech-asr, tabular-forecasting.
- README updated with the multi-kind architecture and new endpoints.
🧪 Tests & tooling
- Go end-to-end suites: import → train → deploy → predict / forecast / embed / rerank, error contracts, metrics propagation (
go test -race ./...). - Python per-task unit tests (
trainer/tests/,python -m pytest). - Docker image installs
trainer[gguf,tabular]extras;.golangci.ymlextended.
⚠️ Upgrade notes
- No breaking changes — existing
causal_lmtasks work as before (migrated to the default kind automatically). - Simulation backend remains the default for every kind (full flow without a GPU). For real training/serving: set
TRAINING_BACKEND=local(andINFERENCE_BACKEND=llamacpp) +pip install -r trainer/requirements.txt. - Optional dependencies (
statsforecast, Chronos,torchao) are guarded, experimental and never required.
📦 New API endpoints (selected)
POST /api/v1/tasks/{taskID}/examples/import-vision-zip | import-vision-pdf | import-asr-zip
POST /api/v1/inference/{id}/embed | /rerank | /embed-corpus | /transcribe | /forecast | /predict-batch
POST /api/v1/tasks/{taskID}/tables (CSV upload for tabular)
POST /api/v1/tasks/{taskID}/training/tabular | /training/forecast
GET /api/v1/tasks/{taskID}/tables/{id}/preview