Modelpublic

BastienHot/buzzwords-style-loras

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes37downloads

Model Card

Buzzwords & Misdemeanors — style LoRAs

Eight GGUF LoRA adapters for `openbmb/MiniCPM5-1B` (apply with llama.cpp, no merge), one per jargon style. They turn the base model into a courtroom actor (judge / prosecutor / defense — selected by system prompt) that argues in dense domain jargon while never naming the hidden profession or charge in plain words.

Built for the Hugging Face Build Small hackathon. See the game: BuildSmallHackathon · base GGUF: openbmb/MiniCPM5-1B-GGUF.

Adapters

style-{corporate, aviation, ai, politics, medical, gaming, sports, scifi}.lora.gguf

How they were trained

—Teacher distillation: Gemma 4 31B (FP8) generated ShareGPT courtroom transcripts, anchored by curated per-style jargon banks and seeded for variety.
—Curriculum fork: stage 1 trains a generic legal LoRA (legal_base) on MiniCPM5-1B; stage 2 forks each style from it (fresh run, same config) — never merged into the base.
—Each adapter benchmarked vs the base (think-mode off): all 8 PASS — high target-jargon coverage, peaking on their own style.

Usage (llama.cpp)

bash

# base + one style adapter
llama-cli -m MiniCPM5-1B-Q4_K_M.gguf --lora style-corporate.lora.gguf -p "..."