Datasets:

---

LocalLLaMA

/

typed-decisions

Like 159

Follow

LocalLLaMA 216

Nirnaya fine-tuned specialist results

Model: kshitijthakkar/nirnaya-0.4b-0.2a-decision
The model card now includes the benchmark result, workflow chart, and Hugging Face .eval_results/typed-decisions.yaml metadata: model PR #1 (merged).

• Mode: fine-tuned specialist. The upstream all/train split has 6,000 rows. Our training manifest records 5,735 in the training split, 135 in dev, and 130 in calibration. all/test was excluded from training. Compare with fitted specialists, not the zero-shot table.
• Evaluation: all/test, 400 cases / 2,000 decisions; zero inference errors.
• Overall accuracy: 0.5075.
• KL(gold || model): 0.32660 nats; Brier: 0.17917; top-label ECE: 0.05083.
• Observability workflow (agent_trace_observability): 0.376 accuracy over 500 decisions. Other workflow accuracies: customer service 0.446, invoice processing 0.594, security incidents 0.614.
• Serving: NVIDIA L4. Each case adapter call receives the shared state and all five questions, then scores the five training-compatible per-question prompts together in one padded model forward pass. The questions are not jointly conditioned on each other. Model load excluded. Case latency p50 45.62 ms, p95 51.26 ms, mean 46.45 ms.
• Model revision: adcca573db021c43718984baa911b47bdfd045df.

The report defines scoring and latency scope and includes per-decision outputs, environment details, and the evaluator. ECE uses 10 equal-width bins and top-label confidence versus argmax agreement. Brier is the sum of squared probability errors over options. These are self-reported community results and have not been independently reproduced by leaderboard maintainers.