AVA
A 2B model fine-tuned to beat Llama 3.2 3B on ARC — with a 42 MB adapter on a 4 GB GPU.
Research/training stack for a tool-using, memory-aware assistant targeting 4 GB VRAM (RTX A2000 laptop). 17-benchmark, 16,872-task eval harness. Custom Triton kernel work, verifier-RL, external memory, published HuggingFace adapter.
| Metric | Value |
|---|---|
| ARC-Challenge | 82.0% |
| ARC-Easy | 92.0% |
| Llama 3.2 3B-Instruct baseline | 78.6% |
| MMLU 5-shot | 59.2% |
| GSM8K (greedy / k=5) | 35.3% / 44.0% |
| Adapter size | 42 MB |
| Training peak VRAM | 1.81 GB |
| Eval harness | 17 benchmarks / 16,872 tasks |
