praveenvnktsh00/minutes-notes-gemma4-e2b

Minutes live notes (Gemma 4 E2B LoRA)

A LoRA for Gemma 4 E2B that keeps a meeting's notes up to date while it is happening. Every 30 seconds of transcript it reads the current notes and the new window, and answers with a few compact edit ops: start a topic, add a note, add a sub-bullet, edit or drop a note, or . for nothing. It runs on a Mac next to live transcription; it is the live notetaker in Minutes.

Since stage 3 the live notes leave action items to the end-of-meeting wrap-up (minutes-tidy-gemma4-e2b): the op grammar has points and decisions only.

Files:

• minutes-notes-lora-v3.gguf: current (stage 3, no action items), the llama.cpp LoRA for a Gemma 4 E2B GGUF (tested on unsloth/gemma-4-E2B-it-GGUF Q4_K_M). SHA-256 5f7d625017112df3c18df7bc1bd5a5ffe3b017946345c59aa6337eba8754d2c9.
• adapters-v3/: the stage-3 MLX-LM adapter.
• minutes-notes-lora.gguf, adapters/: stage 2 (with action items), kept for Minutes versions before 1.5.818.

How to run it

The model must be decoded the way it was evaluated, or it looks broken:

1. Gate the empty answer. At the first output token, answer . only if P(.) >= 0.9; otherwise ban . for that token and decode greedily (max ~300 tokens). Ungated, it answers . almost every window.
2. Repair sub-bullets. A sub-bullet op whose parent is unknown, already has its maximum children, is itself a sub-bullet, or is in a closed topic becomes a plain note.
3. Filter near-duplicates of existing notes and of the user's own notes.

The prompt format (system prompt, the notes state, the window, the user's own notes), the op grammar and the repair rules are in the app: live_summary/finetuned.rs.

Training

• Data: 45 meetings of the AMI Meeting Corpus (CC BY 4.0), split into 30-second windows, with a simulated user typing their own notes. Each window labelled with the ops a careful notetaker would make, by Claude Opus following a written guide, and checked by a rule-based validator (grounded numbers, no narration, note length, topic timing, duplicates).
• LoRA rank 16, scale 2, all layers, MLX-LM, 1.5 epochs, lr 5e-5 cosine, batch 8 (1 x 8 accumulation), max length 1,600 tokens. One Apple Silicon Mac (24 GB), ~2 h.
• Stage 3: the stage-2 labels with every action-item op (104 notes, 16 edits of them) removed and the action kind taken out of the system prompt; trained on an 8-bit MLX copy of the base (the LoRA itself is full precision).

Evaluation

Closed loop on 10 held-out AMI meetings: the model runs a whole meeting on its own notes, as the app does.

	ops	validator op pass	window pass	adds/min
reference labels (Opus)	520
stage 1 (Haiku labels)	477	0.887	0.903	0.57
stage 2 (Opus labels)	440	0.852	0.878	0.65
this model (stage 3, no action items)	408	0.858	0.875	0.55

Stage 3 makes no action-item ops (0 of 408) and finds topics like stage 2 (5.2 per meeting against 6.8 in the reference; stage 1 found 1.9). Known weaknesses: notes run 1-3 words over the 14-word limit (39 of 408 ops), and a topic is sometimes opened too soon after the last (12). Stage 3's closed loop ran on the 8-bit base it was trained on; stages 1-2 ran on the bf16 base.

Credits

AMI Meeting Corpus: Carletta et al., 2005, CC BY 4.0. Base model: Google Gemma 4 E2B.