MingZwhy/Qwen3-0.6B-W2.79A8KV16-OPD-3phase
Text Generation • 0.6B • Updated
Quantization-Aware On-Policy Distillation for Qwen3. QAT starting points + the 3-phase recovered model. Code: github.com/MingZwhy/QAOPD
Note Deployment export -- loads directly with trust_remote_code=True. GSM8K 43.14 / MATH-500 24.20 / AMC23 avg@16 4.69.
Note Latent QAT checkpoint -- the OPD starting point for W2.79. Loading it plainly returns UNQUANTIZED weights.
Note Latent QAT checkpoint for W1.88. Not yet run through OPD.