Datasets:
Submission: apracv/pramana-2b (2B) and apracv/pramana-0.8b (0.8B), fine-tuned on train: 0.8005 / 0.7970 acc, KL 0.0755 / 0.0819
For the fitted / fine-tuned table.
apracv/pramana-2b
Model: apracv/pramana-2b, two LoRA adapters on Qwen3.5-2B whose output distributions are averaged; returns a distribution over every option.
Training: fine-tuned on the train split only (80 train cases held out for temperature fitting). Nothing from test.
Scoring: full test split (2,000 decisions), one forward pass per adapter per request.
| Accuracy | KL from gold | Brier | ECE |
|---|---|---|---|
| 0.8005 | 0.0755 | 0.0415 | 0.1675 |
By type: choice 0.767, noul 0.858, score 0.782.
By workflow: agent_trace_observability 0.750, customer_service 0.830, invoice_processing 0.852, security_incidents 0.770.
apracv/pramana-0.8b
Model: apracv/pramana-0.8b, one LoRA adapter on Qwen3.5-0.8B; returns a distribution over every option.
Training: fine-tuned on the train split only (80 train cases held out for temperature fitting). Nothing from test.
Scoring: full test split (2,000 decisions), one forward pass per request.
| Accuracy | KL from gold | Brier | ECE |
|---|---|---|---|
| 0.7970 | 0.0819 | 0.0446 | 0.1646 |
By type: choice 0.762, noul 0.865, score 0.772.
By workflow: agent_trace_observability 0.744, customer_service 0.822, invoice_processing 0.838, security_incidents 0.784.
Thanks for submitting both. They are now on the card in the fitted / fine-tuned table:
- pramana-2b at #8: 0.8005 accuracy, KL 0.0755, Brier 0.0415, ECE 0.1675
- pramana-0.8b at #10: 0.7970 accuracy, KL 0.0819, Brier 0.0446, ECE 0.1646
The Kind column notes that both are temperature-calibrated on 80 held-out train cases. The post didn't include latency, so that cell shows a dash. If you post a p50 per case and the hardware it ran on, I'll add it.