Datasets:

---

mshojaei77

/

tasmim-bench

Like 3

Hi @mshojaei77 , thank you for building Tasmim-Bench and for including Decima in your first run.

This PR adds screening predictions for three Decima checkpoints under submissions/amyrmahdy-decima/: one Parquet file and the same rows as JSONL, with the columns of versions/v1.1.2/data/predictions_screening.parquet (384 rows per model). I also added run_decima.py, the script that produced them.

I computed the metrics with the definitions in docs/SUBMIT_RESULTS.md. The same code gives your published decima-base row (0.479 / 0.716 / 0.958 / 62.5) from predictions_screening.parquet.

decima-base gets 65.3 here and 62.5 in your run of the same weights. The two runs agree on 88% of labels, so the gap comes from how the questions are passed in (my mapping is described below). I would treat your run as the reference.

Model

• Model ID / provider: amyrmahdy/decima-agent, amyrmahdy/decima-base, amyrmahdy/decima-base-clean (local)
• Parameters and license: 321,125,381 each (mmBERT-base encoder 306,939,648 + scoring head 14,185,733, counted from the safetensors in each repo); Apache-2.0
• Variant, precision, quantization: the onnx/int8 files in each repo (8-bit weight-only MatMulNBits, block 32, for the encoder; dynamic int8 for the scoring head)
• Revision: decima-agent 652d90ac9d33088edd730692c95f1da381cb9174; decima-base e5d84ab3a8436cedc45b0daefa91979f68319455 (the revision you ran, same int8 files as the 2.0 upload); decima-base-clean 35eb218bdeeb8c450baba4159ebc08b67c4d632c (a later commit only changed the card)
• Native support: Choice / Noul / Score all native. Mapping: Decima's TypeSafe-style adapter (decima.systemone.to_question). Choice options are key: description in the published order, Noul is the proposition text with options yes/no (true = P(yes)), Score uses the four published anchors with level 0 first. The state is the published JSON string, unchanged. I used lang="fa", which turns on Decima's Persian script normalization (ی/ک, digits). Predicted label = argmax, so Noul is true when p ≥ 0.5.
• Result type: zero-shot
• Tasmim-Bench data used for training or tuning? No. All three were trained before Tasmim-Bench's first commit (2026-10-13 05:02 UTC). decima-base 2.0 was published on 2026-10-04. The decima-base-clean weights were final on 2026-10-11 and the decima-agent 2.3 weights on 2026-10-12. Both were published on 2026-10-18. No weights, calibration, question wording or options were tuned on Tasmim data.
• Public benchmark exposure before evaluation? The benchmark has been public since 2026-10-13. On 2026-10-18 I ran the 24 development cases once, only to evaluate an unreleased experimental model, with decima-base-clean as one of the reference points. This PR is the first time I ran the screening cases. For decima-base only, I also tried the default lang="en" on screening. It gets 63.6 instead of 65.3. I report fa because the cases are Persian.

Results

• See the table above. Overall* = 100 × mean(Choice Acc, Noul F1, 1 − Score MAE / 3)
• Valid decisions: 192 Choice + 96 Noul + 96 Score = 384/384 for each model
• Prediction file: submissions/amyrmahdy-decima/predictions_screening_decima.parquet (+ .jsonl)

Runtime

• Hardware: ARM64 workstation CPU, CPU only (no GPU used). The machine was running other jobs at the same time, so treat the timings as approximate.
• Runtime and version: ONNX Runtime 1.30.0 (CPUExecutionProvider), Decima runtime 2.3.0 (pip install decima-ai), Python 3.12.3, transformers 5.17.0 (tokenizer only)
• Batch size and calls per case: one decide_many call per case with all 4 questions. Up to 16 rows and 512 padded tokens per encoder call, 4 intra-op threads, 1 inter-op thread.
• Warm-up / timed cases / timing method: 5 development cases as warm-up, then 3 timed passes over the 96 screening cases. Each case's latency is the median of its 3 passes. time.perf_counter() starts before the questions are mapped and tokenized and stops after every probability vector has been parsed and checked. Model download and load are not timed. Decima caches option encodings, and I cleared that cache before each case so no case reuses work from another. P50/P95 use linear interpolation at (n − 1) × p.
• P50 / P95 milliseconds per case: see the table
• Failures or unsupported questions: none

About the earlier latency

Your v1.0.7 latency table lists decima-base at 2,539 ms per case (i7-13650HX, Windows). I think most of the difference is the thread count. Decima.from_pretrained uses 1 thread unless told otherwise, and these checkpoints read the state once per question: 4 passes of about 200 tokens each (state plus question) per case, plus about 12 short option rows. With 1 thread on the same CPU I get 1,537 ms P50 per case (one call per question). That is much closer to your number, and the rest is probably hardware. The numbers above use threads=4. They are different settings, and I don't think either measurement is wrong.

Happy to rerun anything with your harness if you prefer.

A. M. Madani (amyrmahdy)