CoolFace
Modelpublic

vXofi/businessgpt-v16-qwen3.5-9b-gguf

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
0likes70downloads
Model Card

BusinessGPT v16 Qwen3.5 9B GGUF

This repository contains merged GGUF exports of the BusinessGPT v16 LoRA and its Qwen3.5 9B abliterated base for local or CPU inference with llama.cpp.

Files And Evaluation Status

  • —Q5_K_M is the deployed and evaluated production quant.
  • —Q4_K_M is provided as a smaller alternative but has not completed a controlled quality comparison against Q5.

The Q5 configuration was evaluated on 43 private out-of-time text contexts. One primary reviewer judged 51.2% of responses good and 37.2% acceptable, for 88.4% reviewer-usable. This is an ordinal single-reviewer result, not a production acceptance rate.

Additional controlled results:

ComparisonResult95% session-cluster CI
v16 artifact vs unadapted base artifact, 40 contextsv16 preference 80.0%67.9%-88.9%
v15 vs v16, 64 contextsv15 preference 61.7%49.3%-73.3%
production vs legacy prompt, 43 contextsproduction preference 55.8%45.0%-65.3%

See the public evaluation report for methodology, label definitions, experiment caveats, and limitations.

Intended Use

The model is intended for research and private, allowlisted Russian informal group-chat deployments. It is not a general assistant, factual authority, moderation model, or safety-critical system.

Risks

The persona permits strong profanity and provocative humor, and the base model is abliterated. Outputs can be offensive, discriminatory, sexual, biased, factually incorrect, or otherwise unsafe. Do not expose the model through an unrestricted public endpoint.

Raw training conversations and evaluation artifacts are private. Memorization risk is reduced through filtering and access-controlled deployment, not eliminated. Vision quality, Q4/Q5 equivalence, long multi-turn stability, and inter-rater agreement have not been established.

Read the complete responsible-use statement before deployment.

Links