Hugging Face's logo

Datasets:
LocalLLaMA
/
typed-decisions

Benchmark
Tasks:
Text Classification
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
structured-decisions
calibration
probabilistic-classification
system-one
workflow-evaluation
Synthetic
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
53

Add Sagicc/nanoGentzen-Jev-v3 (4.93M params) — 0.731 acc, 0.150 KL, 0.081 Brier, 0.087 ECE

#52
by Sagicc - opened
Discussion
Sagicc

Hi! Here are the test split evaluation results (400 cases / 2,000 decisions) for Sagicc/nanoGentzen-Jev-v3 (4.93M parameters).

The .eval_results/Sagicc__nanoGentzen-Jev-v3.yaml file and full prediction traces have been merged into the model repository.

Summary Metrics (test split — 400 cases, 2,000 decisions)

Model Params Mode / Kind Accuracy ↑ KL from gold ↓ Brier ↓ ECE ↓ p50 latency
Sagicc/nanoGentzen-Jev-v3 4.93M general, open weights (MIT) 0.7310 0.1498 0.0809 0.0871 69.4 ms

Breakdown by Question Type

  • noul Accuracy: 0.8283
  • choice Accuracy: 0.6850
  • score Accuracy: 0.6925 (score ±1: 0.9725, score MAE: 0.2843)
  • Protocol: One request per case (state + all five questions), 0 errors, total wall time 40.25s.

Breakdown by Workflow (500 decisions each)

Workflow Accuracy ↑ KL from gold ↓ Brier ↓ ECE ↓ noul Acc choice Acc score Acc score ±1
invoice_processing 0.8460 0.0876 0.0503 0.1183 0.9450 0.7300 0.8050 0.9800
security_incidents 0.7280 0.1010 0.0573 0.1192 0.7800 0.7400 0.6700 0.9800
agent_trace_observability 0.7060 0.0970 0.0584 0.1237 0.8400 0.6550 0.6900 1.0000
customer_service 0.6440 0.3135 0.1576 0.0504 0.6800 0.6650 0.6050 0.9300

Reproducibility & Summary JSON

Served locally via nanoGentzen-Jev (LocalGentzenJevEngine, Stage-3 Span Pointer Head + 3-Class Intuitionistic Epistemic Value Head + 2-Hop Horn Grounding) exposing the /v1/systemone and /v1/decide endpoints.

{
  "model": "Sagicc/nanoGentzen-Jev-v3",
  "n_cases": 400,
  "n_decisions": 2000,
  "accuracy": 0.731,
  "kl_from_gold": 0.1498,
  "brier": 0.0809,
  "ece": 0.0871,
  "accuracy_noul": 0.8283,
  "accuracy_choice": 0.685,
  "accuracy_score": 0.6925,
  "score_within_1": 0.9725,
  "score_mae": 0.2843,
  "p50_latency_ms": 69.38,
  "wall_time_s": 40.25
}
sandeep85
 LocalLLaMA org

Thanks for the run! I've added nanoGentzen-Jev-v3 to the zero-shot table at #15 (0.731 acc, 0.150 KL, 0.081 Brier, 0.087 ECE, 69 ms on your hardware) and to the chart. It's in PR #53 until it's merged.

Your eval file looks good (same shape as d3-turbo's, all four task ids match the dataset's eval.yaml, and your model page already shows the four scores). It just isn't in the dataset's leaderboard list yet, maybe the Hub catching up. If it's still missing in a day or so, a small commit touching the YAML is worth a try.

One question: what was it trained on? The zero-shot table is for models that have never seen these workflows or their question schemas. If any of the typed-decisions train split went in, I'll move it to the fitted table.
Sagicc

Thank you! This dataset is used for validation only, so zero-shot is real place:). Here is the base model https://huggingface.co/Sagicc/nanoGentzen-v2 (with train code and dataset there). Looking how to improve the architecture further, specifically for the Jev :))

Sign up or log in to comment