AutoAEG checkpoints

Minimal PEFT checkpoint assets for the current AutoAEG experimental chain:

checkpoints/sft10k_v3_final/: final SFT adapter trained from scratch on sft10k_v3.
checkpoints/rms_span_final/: final improved structured-IoU + RMS/span GRPO adapter initialized from that SFT.
checkpoints/cls_span_latest/: latest improved structured-IoU + CLS/span GRPO adapter, global step 1300, initialized from the same SFT. training_aux/reference_lora.pt is the fixed KL reference adapter needed to continue GRPO; training_aux/class_templates.pkl is the CLS reward template cache.

The base Qwen model is not included. Load the corresponding base model first, then attach the selected PEFT adapter. Optimizer, scheduler, RNG, and distributed-training shards are intentionally excluded. Exact source paths, steps, sizes, SHA-256 digests, and tensor counts are recorded in CHECKPOINT_MANIFEST.json.

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support