AdamProAI/eyebench-v3-qwen35-4b-lora-smoke-verify

LoRA adapter exported from Tinker for EyeBench-style visual reasoning experiments.

Base Model

• Qwen/Qwen3.5-4B

Verification Snapshot

• Benchmark score: 40.00%
• Exact match: 2/5
• MAE: 5.40

Notes

• This is an exported adapter checkpoint, not a merged full model.
• The verification run used a tiny smoke split, so the score is only a pipeline sanity check.