VON-3B
This is a hackathon submission repo and my 3B model. designed and ran the full path: SFT, RL, group-conditioned adaptive LoPD (rl & distillation method from my research), LoRA, weight edits, and the laptop GGUF pack.
VON-3B is one 3B model that covers two jobs: an offline coding assistant and an autonomous agent. after one public download it runs locally in llama.cpp or any inference engine on a standard 8 GB machine. no API key / network at inference.
we built it to:
- write and repair code
- keep reasoning short (base reasoned too long and took lots of time)
- emit a real one-line tool call (
<tool_call>{...}</tool_call>) so it can act as an agent, not only a chatbot
other details:
- model card and weights: https://huggingface.co/josephmayo/von3b
- laptop artifact:
von3b-Q8_0.gguf (llama.cpp, GGUF Q8_0, 3,285,475,488 bytes)
- runtime: llama.cpp only
- target machine: 4 vCPU, 8 GB RAM, integrated GPU, Ubuntu 22.04
- writeup:
REPORT.md
we started from WeiboAI/VibeThinker-3B. On a matched EvalPlus 0.3.1 HumanEval check (same 164 tasks, greedy max_new=8192), the model beats that snapshot:
| Arm | HumanEval pass@1 | HumanEval+ pass@1 |
|---|
| VON-3B | 0.921 (151 / 164) | 0.884 (145 / 164) |
| VibeThinker-3B base | 0.866 (142 / 164) | 0.817 (134 / 164) |
we report that comparison because we were compute-constrained - thats why we didnt run more evals. tool probe, 32 tasks, greedy max_new=256, same snapshot: the model emits a valid one-line <tool_call> with short think on 32 / 32. The base emits 0 / 32.
download:
1bash download_model.sh
2llama-cli -m model/von3b-Q8_0.gguf -c 65536 -ctk q4_0 -ctv q4_0 -ngl 0
ctx is 65,536 with Q4_0 K/V cache so the 8 GB profile can hold long coding sessions.
required package files (official template):
metadata.jsondownload_model.shREPORT.mdmodel/ (GGUF downloaded by the script, not committed)
