Tier	What you get
Free	Core pipeline, basic AI assist, local LLM only
Paid	Cloud LLM, integrations, full AI generation suite
Premium	Fine-tuned models, multi-user, advanced analytics
Ultra	Human-in-the-loop operator

License

Discovery pipeline: MIT · AI features: BSL 1.1 (→ MIT after 4 years)

## [AVCT] Avocet > Email classifier training tool — multi-account IMAP, card-stack labeling ### Repository https://git.opensourcesolarpunk.com/Circuit-Forge/avocet ### Milestones https://git.opensourcesolarpunk.com/Circuit-Forge/avocet/milestones ### Design Pillars Privacy · Safety · Accessibility — co-equal, non-negotiable. Neurodivergent and adaptive needs users are a primary audience. ### Tier System | Tier | What you get | |------|--------------| | Free | Core pipeline, basic AI assist, local LLM only | | Paid | Cloud LLM, integrations, full AI generation suite | | Premium | Fine-tuned models, multi-user, advanced analytics | | Ultra | Human-in-the-loop operator | ### License Discovery pipeline: MIT · AI features: BSL 1.1 (→ MIT after 4 years)

pyr0ball added the

priority:medium

status:concept

labels 2026-03-10 20:26:24 -07:00

pyr0ball commented

2026-03-15 16:20:51 -07:00

Author

Owner

Fine-tune classifier pipeline — completed 2026-03-15

Branch feat/vue-label-tab in avocet repo now includes the full fine-tune training pipeline:

New capabilities:

scripts/finetune_classifier.py — fine-tunes deberta-small (100M) or bge-m3 (600M) on labeled JSONL data
Multi-file JSONL support: accumulate multiple labeling sessions, last-write-wins dedup by content hash
FineTunedAdapter — auto-discovered at benchmark startup via models/*/training_info.json
GET /api/finetune/run — SSE-streamed training progress; GET /api/finetune/status — lists trained models
BenchmarkView.vue — trained models badge row + collapsible fine-tune section with auto-benchmark on completion

Test coverage: 95 new tests (15 finetune unit, 5 adapter, 4 benchmark, 6 API, 1 integration)

See Circuit-Forge/avocet#8 for full implementation notes.

## Fine-tune classifier pipeline — completed 2026-03-15 Branch `feat/vue-label-tab` in `avocet` repo now includes the full fine-tune training pipeline: **New capabilities:** - `scripts/finetune_classifier.py` — fine-tunes `deberta-small` (100M) or `bge-m3` (600M) on labeled JSONL data - Multi-file JSONL support: accumulate multiple labeling sessions, last-write-wins dedup by content hash - `FineTunedAdapter` — auto-discovered at benchmark startup via `models/*/training_info.json` - `GET /api/finetune/run` — SSE-streamed training progress; `GET /api/finetune/status` — lists trained models - BenchmarkView.vue — trained models badge row + collapsible fine-tune section with auto-benchmark on completion **Test coverage:** 95 new tests (15 finetune unit, 5 adapter, 4 benchmark, 6 API, 1 integration) See Circuit-Forge/avocet#8 for full implementation notes.

pyr0ball added

status:active-dev

and removed

status:concept

labels 2026-03-29 20:05:56 -07:00