🗡️ Rostam R1 Stage 13-2 (DPO Aligned)
Rostam R1 Stage 13-2 is the Direct Preference Optimization (DPO) aligned release of the Rostam series, fine-tuned from prog-love/rostam-r1-stage13-1.
Verified Benchmark Results (PQuAD Test - 8,002 samples):
- Exact Match (EM): 42.08% (+4.89% over Stage 12, +10.59% over Stage 13-1 raw).
- F1 Score: 67.20% (+3.83% over Stage 12).
- False Abstention Rate: 1.23% (Down from 7.24% in Stage 12).
- Repetition Mitigation: Enforced
<|turn|>stop boundaries, eliminating repetition loops.
Training Setup:
- Lineage:
stage12-part2->stage13->stage13-1->stage13-2 - Method: Full-Parameter BF16 DPO (all 5.104B parameters trained)
- Optimizer: Native PyTorch Fused AdamW
- Hardware: 1x NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB VRAM)