MiniCPM5-1B-CURSOR — agentic meeting summarizer (sub-1B on-device)

Fine-tune of openbmb/MiniCPM5-1B on the CURSOR protocol: a streaming agent that converts long meeting transcripts (zh-TW / en) into structured, fully-anchored meeting notes by emitting edit ops (ADD/UPD/DEL/CMP/NOP) against one evolving NOTES state — no map-reduce, no ReAct loops.

Try it live: cursor-wasm-demo runs this model entirely in your browser via WASM — watch the NOTES state build chunk by chunk, no server involved.

Checkpoints in this repo

file	status	notes
minicpm5-1b-cursor.Q4_K_M.gguf	locked main (p15d)	the deployed checkpoint — see results below
minicpm5-1b-cursor-p20.Q4_K_M.gguf	candidate, not yet locked	real-ASR-majority retrain, see below
minicpm5-1b-cursor-p19c.Q4_K_M.gguf	superseded experiment	coverage-focused dose, worse on the published T1 numbers than p15d
minicpm5-1b-cursor-p13.Q4_K_M.gguf	superseded	earlier round, kept for the historical record

How p15d (locked main) was made

• Base: openbmb/MiniCPM5-1B (4k context, linear+full attention hybrid).
• Data: teacher traces from Gemma-4-31B-it (NVFP4, thinking ON), judge-filtered against a local gpt-oss-20b judge (only judge-verifiable teacher bullets become SFT targets), 100% synthetic transcripts (zh-TW + en, revision-dense meetings by design).
• Guards: two deterministic harness guards run at render time — promote_decision_summaries (decision-shaped SUMMARY bullets move into DECISIONS) and enforce_decision_chain (opposing-polarity bullets on one subject: keep the latest).
• Verifier: paired with Luigi/granite-4.0-350m-verifier (Apache-2.0, ~215MB), which gates DECISIONS/ACTIONS ops in-stream and re-verifies every final bullet against whole-transcript evidence. The verifier is required for the 0-inversion deployment number below — the model alone measures 4/20.

Measured results (T1 tier, n=20, local judges, 3x majority)

configuration	INVERT	FAITH	COVER	SYNTH
p15d, model-only (raw)	4/20	3.57	3.00	2.30
p15d + verifier, in-stream + sweep (deployed)	0-2/20	4.43	2.85-3.80	2.35-3.40
map-reduce baseline (Qwen3.5-9B)	3/20	3.50	3.05	2.60

G1 capability screen: PASS both languages (with the guards). Verifier agreement with gpt-oss-20b: 97% en / 92% zh on clean evidence (zh discrimination on real-ASR-noisy evidence remains a known, documented gap — see the agentic-summarizer repo's RESULTS.md for the full retraining history and every measured negative on that front).

p20 (real-ASR-majority candidate, 2026-08-18): measured positive, not yet locked

p15d was trained on 100% synthetic transcripts. p20 continues from the p15d checkpoint on the first genuinely real-ASR-majority mix (58.9% real / 41.1% clean-synthetic) — 23 real zh-TW-majority podcast transcripts (donated for this project, teacher-traced and judge-filtered the same way as the synthetic data), oversampled against a clean-synthetic minority.

Held-out real-meeting comparison (3 whole meetings withheld from training, never seen by either checkpoint during training):

meeting	lang	p15d valid-op	p20 valid-op
q08	zh-TW	80%	88%
ho4	zh-TW	100%	100% (tie)
p02	en	86%	100%

p20 improves or ties valid-op on every held-out real meeting, and in spot checks avoided a fabrication p15d made on one of them (p15d invented a DECISIONS bullet — "Use Linux as the preferred operating system" — with no grounding in the source, a job-interview-prep podcast).

Caveat, stated plainly: n=3 held-out meetings is directional, not a ship-gate measurement. p20 is published here as a measured, honest candidate — not a claim that it now supersedes p15d as the recommended deployment. Promoting it would need the same T1-tier (n=20) confirmation p15d itself was held to. Use p15d for anything you need a locked answer on; try p20 if your workload is real-ASR-majority zh-TW/en content specifically and you want to see the newer checkpoint's behavior on it.

Usage

llama.cpp server, greedy, thinking OFF (the model emits hybrid <think> unless disabled):

llama-server -m minicpm5-1b-cursor.Q4_K_M.gguf --n-gpu-layers 999 --ctx-size 4096 \
  --parallel 1 --flash-attn on --jinja --reasoning off --temp 0

The client is the CURSOR harness (agentic-summarizer repo, eval/run_arms.py --arms cursor --base-url <server> --tokenizer openbmb/MiniCPM5-1B --budget 2048). The model expects the exact system prompt and NOTES rendering the harness produces — it is NOT a general chat model.

Caveats (must accompany every reported number)

• zh T2 tier is synthetic; contested-zh (real, decision-bearing zh-TW meetings, as opposed to podcasts) remains the largest unmeasured gap in this project.
• Judge-noise floor +-0.4-0.5 (FAITH/SYNTH); n=20 per tier; reduced cells directional.
• On-device envelope: ~650 MB Q4_K_M, 4k context, per-step ~2.9k prefill tokens.
• p20's held-out comparison is n=3 meetings — directional, not a ship-gate result.