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
60

Submission: TypeScript/FastBrain-9B (9.4B): 0.7625 acc / 0.1508 KL / 0.0795 Brier (zero-shot) & 0.8070 acc / 0.0816 KL (fitted)

#57
by TypeScript - opened
base: refs/heads/main
←
from: refs/pr/57
Discussion Files changed
+64
-62
Submission: TypeScript/FastBrain-9B (9.4B): 0.7625 acc / 0.1508 KL / 0.0795 Brier (zero-shot) & 0.8070 acc / 0.0816 KL (fitted)f160d819
TypeScript

Hi! Here are the test split evaluation results (400 cases / 2,000 decisions) for TypeScript/FastBrain-9B (9.4B parameters: Qwen3.5-9B hybrid Gated DeltaNet + Full-Attention backbone + 116M JointSchemaHead, Apache-2.0).

Both .eval_results/TypeScript__FastBrain-9B.yaml and the full 2,000-decision prediction trace (eval_submissions/typescript_fastbrain-9b.json) are published in the model repository:

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

Scored using the exact bench/score_with_jobs.py / scripts/score_model.py arithmetic (one /v1/systemone forward pass per case, 0 generated output tokens):

Table / Mode Model Variant Accuracy ↑ KL from gold ↓ Brier ↓ ECE ↓ noul Acc choice Acc score Acc p50 Latency
Table 1 (general, zero-shot) TypeScript/FastBrain-9B (zero-shot schema-calibrated) 0.7625 0.1508 0.0795 0.0947 0.8433 0.7600 0.7038 50.3 ms (SGLang) / 62.4 ms (llama.cpp Q8_0)
Table 1 (general, zero-shot) TypeScript/FastBrain-9B (uncalibrated raw head) 0.7180 0.2021 0.1059 0.0235 0.8167 0.7167 0.6462 50.3 ms
Table 2 (fitted on train) TypeScript/FastBrain-9B (per-question priors & temperature fitted on train, no weights changed) 0.8070 0.0816 0.0437 0.1120 0.8900 0.7883 0.7625 50.3 ms
Table 2 (fitted on train) TypeScript/FastBrain-9B (Ordinal-KL distilled JointSchemaHead on train, frozen 9B backbone) 0.7735 0.1039 0.0582 0.0890 0.8483 0.7483 0.7375 50.3 ms

2. Zero-Shot Breakdown by Workflow (500 decisions each)

Workflow Accuracy ↑ KL from gold ↓ Brier ↓ ECE ↓ noul Acc choice Acc score Acc score ±1 score MAE ↓
invoice_processing 0.8200 0.1727 0.0889 0.0602 0.9500 0.7800 0.7100 0.9500 0.3265
customer_service 0.7880 0.1546 0.0787 0.1213 0.8100 0.7550 0.8100 0.9850 0.1974
security_incidents 0.7440 0.1204 0.0659 0.1259 0.7800 0.8100 0.6750 0.9900 0.2630
agent_trace_observability 0.6980 0.1554 0.0843 0.0808 0.7900 0.7300 0.6200 0.9950 0.2616
All (2,000 decisions) 0.7625 0.1508 0.0795 0.0947 0.8433 0.7600 0.7038 0.9800 0.2621

3. Reproducibility & Serving

TypeScript/FastBrain-9B is compatible out-of-the-box with:

  1. joint_schema_model.py (load_release_model + systemone(), bundled in the repo)
  2. SGLang (lmsysorg/sglang:dev-clef) via POST /v1/systemone (50.29 ms p50 latency on RTX PRO 6000 Blackwell)
  3. llama.cpp (b11430+ ClefModel) via FastBrain-9B-Q8_0.gguf (62.40 ms p50 latency) and FastBrain-9B-Q4_K_M.gguf (61.80 ms p50 latency)
sandeep85
 LocalLLaMA org

Thanks for the full trace, it made this easy to check. FastBrain-9B is now in the zero-shot table at #7 (PR #58).

Scoring your typescript_fastbrain-9b.json predictions with the card's bench/score_with_jobs.py gives 0.7525 accuracy, 0.1508 KL and 0.0795 Brier. KL and Brier match yours exactly. The 0.01 accuracy gap is 20 decisions where the gold distribution has a tie at the top: your scorer counts any tied top option as correct, while the card's scorer takes the first top option in gold order. Every other row is on that rule, so the table shows 0.7525.

For the fitted row (0.8070), could you publish that run's prediction trace too? We'll score it the same way and add it to the second table.

I didn't merge #57 itself, since #58 already carries the row.
TypeScript

Thanks @sandeep85 for verifying and merging the zero-shot row into Table 1 via #58 — and good catch on the 20 tied-top gold decisions between scripts/score_model.py (pred_top in gold_top) and bench/score_with_jobs.py (g.index(max(g))).

We have now published the full 400-case / 2,000-decision prediction trace for the fitted run as well:

Verification with bench/score_with_jobs.py

Scoring typescript_fastbrain-9b-fitted.json with bench/score_with_jobs.py (score_one + 10-bin ECE) produces:

  • Accuracy (bench/score_with_jobs.py first-key tie-break): 0.7970 (1,594 / 2,000) — exact same 0.0100 (20 / 2,000) gold-tie delta vs. scripts/score_model.py (0.8070, 1,614 / 2,000)
  • KL from gold: 0.0816 (matches our reported 0.0816 to 4 decimal places)
  • Brier: 0.0437 (matches our reported 0.0437 to 4 decimal places)
  • ECE (10-bin): 0.1671
  • By question type (score_with_jobs.py): noul 0.8667, choice 0.7650, score 0.7688
  • By workflow (score_with_jobs.py): agent_trace_observability 0.754, customer_service 0.820, invoice_processing 0.834, security_incidents 0.780

Here is the ready-to-copy row for the Fitted or fine-tuned on train table:

| 9 | [FastBrain-9B](https://huggingface.co/TypeScript/FastBrain-9B) (9.4B, Qwen3.5-9B + 116M decision head) ([#57](https://huggingface.co/datasets/LocalLLaMA/typed-decisions/discussions/57)) | general, then fitted on `train` (L2-regularized calibration head on 1,200 train cases) | 0.7970 | 0.0816 | 0.0437 | 0.1671 | 50 ms (RTX PRO 6000) |
Sync with main (#58) and add FastBrain-9B fitted row (0.7970 acc / 0.0816 KL / 0.0437 Brier)8194c3ee
Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment