Datasets:

---

LocalLLaMA

/

typed-extraction

Like 3

Follow

LocalLLaMA 240

Hi, thanks for building this benchmark! Here is a zero-shot submission for an open 150M extractor:

Model: tasksource/UltraBERT-150M-clairic-extraction (150M encoder, Apache-2.0)
Mode: zero-shot. It was never trained on typed-extraction train or test.

score.py output (unmodified scorer, all 1,690 test cases predicted):

{
 "cases_predicted": 1690,
 "cases": 1690,
 "macro_family_exact": 0.6159,
 "macro_family_exact_present": 0.5479,
 "hallucination": 0.1409,
 "miss": 0.2969,
 "exact": 0.591,
 "partial": 0.6107,
 "p_null_auroc": 0.815,
 "ece": 0.4198,
 "exact_at_90_coverage": 0.6214,
 "aurc": 0.2819
}
bank_statement         n=  608  exact 0.546  hallucination 0.021
claim_email            n=  617  exact 0.515  hallucination 0.391
cuad                   n= 4182  exact 0.700  hallucination 0.082
drop                   n= 1684  exact 0.194  hallucination None
extractbench           n= 2115  exact 0.242  hallucination 0.173
invoice                n=  681  exact 0.667  hallucination 0.514
itinerary              n=  608  exact 0.809  hallucination 0.084
job_posting            n=  609  exact 0.678  hallucination 0.394
lease                  n=  770  exact 0.764  hallucination 0.131
medical_appointment    n=  760  exact 0.603  hallucination 0.064
meeting_notes          n=  659  exact 0.569  hallucination 0.037
real_estate_listing    n=  961  exact 0.664  hallucination 0.053
shipping_notice        n=  607  exact 0.741  hallucination 0.426
squad2                 n= 1908  exact 0.747  hallucination 0.267
wikipedia              n=  964  exact 0.800  hallucination 0.024

Predictions: eval/typed-extraction-test-predictions.jsonl in the model repo, so you can re-score them.

How the predictions were made

• One forward pass per field: the field's instructions paired with the document. The model tags the answer tokens (B/I/O), and long documents are read in overlapping 2,048-token windows.
• No span means null, and p_null = 1 − the best span's B probability.
• Spans are mapped to types with plain code:
  • text as written;
  • number/duration: the first number in the span (k/m suffixes expanded);
  • date/time/datetime through dateutil, years defaulting to reference_date;
  • lists item by item.
• No LLM, no prompt, no per-family rules.

Caveats

• It's an extractor, so computed values (totals, counts, day spans) are mostly out of reach (0.014). That is where DROP, and part of the constructed families, lose points. Relative dates are only resolved when dateutil can parse them.
• confidence is the raw span probability and is not calibrated (ECE 0.42). Ranking is fine (p_null AUROC 0.815).
• Contamination check: our training data, the extraction config of tasksource-jev-typed-decisions, includes the public train splits of SQuAD 2.0, DROP and CUAD. Under a 120-character passage match, 23 of the 1,690 test documents (CUAD, DROP, Wikipedia) share text with it. Without those 23 cases, the headline is 0.616 (hallucination 0.142, miss 0.295), so they don't drive the score.

Latency: about 1.5 s per case on a shared A30 GPU, unoptimized and dominated by the long CUAD contracts (fields are run one by one). An int8 ONNX export runs in the browser: demo.