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
66

Submission: gnumanth/nullspan v1.3.0 (395M 8K tree-isolated encoder, specialist): 0.7900 acc / 0.0760 KL / 0.0420 Brier / 0.1665 ECE

#55
by gnumanth - opened
Discussion
gnumanth
•
edited 3 days ago

For the fitted / specialist table (Table 2, mode: specialist).

  • Model: gnumanth/nullspan (revision 51acf9e097ade3fb1068b3511490ef2756485927, GitHub tag v1.3.0 / commit de900ac4801c9d9f4b55abff68b4c593123d4ed8, MIT) — pure non-LLM 395M 8K bidirectional ModernBERT-large (NullspanSystemOneEngine) packing the full raw case + all 5 questions and candidate options into a single forward pass per case (0 generated output tokens) with Exact Tree-Isolated Parallel-Option Masking (0.0 cross-option RoPE bleed, 100% option-order & A/B/C-label permutation invariant), 2-layer cross-option + cross-question workflow competition, and 24 auxiliary latent workflow factor heads.
  • Training & Contamination Disclosure (mode: specialist): Stage 1 general System One pretraining (BoolQ, ARC-Challenge, SciQ train splits + programmatic numeric/policy/negation/adequacy rubrics) followed by Stage 2 10-model 5-fold Cross-Backbone OOF Self-Distillation (5x ModernBERT-large + 5x Ettin-400M, OOF train 0.8012 acc / 0.0832 KL) + Trajectory-SWA on the LocalLLaMA/typed-decisions train split (1,200 cases). Not zero-shot. The 400-case test split (2,000 decisions) was strictly untouched during training, architecture selection, and OOF temperature calibration. (Prior v1.2.1 435M DeBERTa-v3 checkpoint remains immutably preserved at 43cd3a296bbc858189d1e183eaed6e89f95e5dfd.)

Official test Split Metrics (400 cases · 2,000 decisions · NVIDIA L4 GPU)

Metric gnumanth/nullspan (v1.3.0) v1.2.1 (prior) Delta
Accuracy 0.7900 (79.00%) 0.7540 +3.60%
KL from gold 0.0760 0.1515 -49.8%
Brier 0.0420 0.0770 -45.5%
ECE (10-bin) 0.1665 0.0944 +0.0721
Latency (NVIDIA L4) 47.82 ms/case (9.56 ms/decision) 390.02 ms/case 8.2x faster
Output Tokens 0 0 0

Breakdown by Workflow & Decision Type (test)

  • By Workflow Accuracy:
    • invoice_processing: 0.8440 (84.4%)
    • customer_service: 0.8140 (81.4%)
    • security_incidents: 0.7620 (76.2%)
    • agent_trace_observability: 0.7400 (74.0%)
  • By Question Type Accuracy:
    • noul: 0.8417 (84.2%)
    • choice: 0.7683 (76.8%)
    • score: 0.7675 (76.8%)

Reproducibility & Verification Artifacts

gnumanth changed discussion title from Submission: gnumanth/nullspan v1.2.1 (435M cross-encoder, general + fine-tuned on train): 0.7540 acc / 0.1515 KL / 0.0770 Brier / 0.0944 ECE to Submission: gnumanth/nullspan v1.3.0 (395M 8K tree-isolated encoder, specialist): 0.7900 acc / 0.0760 KL / 0.0420 Brier / 0.1665 ECE
gnumanth

Updated submission to gnumanth/nullspan v1.3.0 (revision 51acf9e097ade3fb1068b3511490ef2756485927, 395M 8K Tree-Isolated Parallel-Span ModernBERT-large, 1 forward pass per case, 47.82 ms/case on L4): 0.7900 accuracy / 0.0760 KL / 0.0420 Brier / 0.1665 ECE across all 2,000 test decisions (strictly zero test leakage; full predictions uploaded at predictions_typed_decisions_test.jsonl). Prior v1.2.1 revision 43cd3a296bbc858189d1e183eaed6e89f95e5dfd remains preserved.
sandeep85
 LocalLLaMA org

Thanks, nullspan v1.3.0 is on the card in the fitted table at #14: 0.790 accuracy, KL 0.076, Brier 0.042 and ECE 0.1665 (10-bin), at 48 ms per case on an L4. Going from the 435M DeBERTa to a 395M ModernBERT and roughly halving KL in one forward pass per case is a good trade.

Sign up or log in to comment