margin-pilot1-gguf

Margin Pilot 1 — Q4_K_M

An experimental merged SFT-LoRA model for MeetingIntro Margin, derived from empero-ai/Qwen3.8-4B-Distill revision c83cb7aa2999d2f35c43e9ae0634a30eb8985a1e.

This package is Pilot 1, not the later DPO adapter. It is a community Empero/Qwen3.5-architecture derivative, not an official Qwen release.

Evaluation limits

The original developer suite reported 279/288 passing model requests, with three repetitions per fixture. Known failures remain. A separate thinking-disabled workstation subset passed 42/60 using evaluator v3. Those are different evaluation conditions and are not interchangeable. Native validation on a 24 GB Mac passed 76/78 cases; the two misses were an unnecessary tentative-RSVP refusal and a date move missing its destination. The eight-holiday all-day batch passed. All 45 roadmap guards and 22 managed-setup guards passed. A clean managed installation using the published, pinned model succeeded on 2026-09-13 UTC. Median native latency was 4.04 seconds, p95 15.77 seconds, with concurrent build activity; this is not a 16 GB hardware qualification. No claim of 99% real-world reliability is made.

Tools must enforce granted roots, preview and confirmation, reversible operations, and independent memory provenance checks. This model does not confer permission to execute its output.

Reproducibility

Quantized size: 2,783,446,336 bytes. SHA-256 is in SHA256SUMS. Use Ollama with an 8192-token context and temperature 0. Runtime and thinking mode materially affect results; qualify the exact app/runtime combination before deployment.

The model contains merged adapter weights. Training data, personal files, credentials, and user conversations are not included in this distribution package. No guarantee is made that a trained model cannot reproduce training examples.

Changes and attribution

Upstream model by Empero AI, based on Qwen3.5 architecture. Modified with Margin-specific SFT LoRA training, merged to canonical tensor names, and quantized to Q4_K_M. Auxiliary MTP tensors preserved from the pinned source. Licensed under Apache 2.0; see LICENSE and upstream attribution.