pepper-desk-e2b โ€” the MNN singularity desk, distilled

A 2B-class specialist that reads a wire of headlines and writes a grounded broadcast report โ€” sources attributed, rumors adjudicated, unknowns said out loud. It is the research-desk brain of Pepper, the open-source on-device news anchor (pepper.software ยท github.com/bunnycompany/pepper ยท watch her: pepper.watch).

MoltBench (12 bundles, 12 blind judges, deterministic rotation)

Model Blind wins Grounding Adjudication Persona
pepper-desk-e2b (this, 2B) 11/12 88.6% 4.67/5 3.75/5
Qwen2.5-7B-Instruct-4bit 1/12 77.5% 2.67 1.42
pepper-7b (persona LoRA) 0/12 50.0% 2.00 1.83

Benchmark, bundles, scorer, and protocol: bench/ in the repo. The origin story matters: the first Pepper model failed this benchmark against its own base (38.1% vs 64.5% grounding) โ€” that failure became the release gate this model had to clear.

Format: think, then speak

Trained think-then-speak. Given wire notes, she emits DESK NOTES: (a private source-weighing analysis) then ON AIR: (the broadcast). Consumers show or strip the notes; score only the broadcast.

System prompt and wire format: see bench/README.md and the repo's gen_eval_v2 harness. Use max_tokens โ‰ฅ 500 โ€” tighter caps truncate her sign-offs (it cost her one judged bundle).

Training

  • Base: google/gemma-4-e2b-it via mlx-community/gemma-4-e2b-it-4bit
  • LoRA (mlx-lm 0.31, git), lr 4e-5, batch 4, seq 1800, grad-checkpoint, iteration-200 checkpoint selected by validation loss (1.557; later checkpoints overfit โ€” the full curve is documented in the repo)
  • Data: 452 examples โ€” 412 claim-verified wireโ†’report pairs authored against the live August-2026 news cycle with explicit DESK NOTES reasoning (including deliberately contaminated bundles with exemplar adjudications), her 65 real broadcast segments, and a 40-riff persona sprinkle
  • Trained on an M3 Ultra in ~90 minutes; reproducible on consumer Apple Silicon

Limitations

She is grounded, not omniscient: judges recorded occasional invented connective detail, "peer-reviewed" applied to preprints, and style drift on very thin wires. She is built to work FROM provided wire notes โ€” as a freestanding chatbot she is out of her element and says so less often than she should. English-first. Not for advice of any kind.

License & lineage

Weights are a derivative of Gemma and ship under the Gemma Terms of Use. The surrounding desk (app, bench, pipeline) is AGPL-3.0. Lineage: Danger Ghost (VTuber era) โ†’ MNN research anchor โ†’ this desk brain.

MNN โ€” all your models, all the time. ๐ŸŒถ

Downloads last month
19
Safetensors
Model size
5B params
Tensor type
U32
ยท
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support