Tasmim-Bench

Tasmim-Bench evaluates structured Persian decision-making with typed Choice, Noul (binary proposition), and Score questions. It compares local checkpoints and hosted decision APIs on shared Persian and mixed-language scenarios.

Leaderboard

The table includes every model with measured results and the community Mizan submissions. Parameter counts come from pinned local checkpoint audits or the linked parameter metadata; hosted counts are safetensors-index totals and exclude separately stored decision-head files. Approximate counts use ≈, and undisclosed hosted sizes say Undisclosed. — means unsupported, Not measured means supported but not measured, and N/A — incompatible rubric identifies a score scale that cannot be compared with this benchmark. † marks Noul measured through a binary Choice adapter, not a native Noul head. Latency P50 is end-to-end milliseconds per case; its source, question count, P95, hardware, and raw samples are linked below. Local CPU, hosted API, and community GPU timing runs use different hardware and are not directly comparable. Latency is not part of Overall*. Overall* is shown only for complete native three-primitive results and the community-reported Mizan submissions; its formula is 100 × mean(Choice accuracy, Noul true-class F1, 1 − Score MAE / 3), a pilot comparison aid rather than validated utility. Rows with numeric Overall are sorted descending; partial rows follow. Coverage and submission provenance are linked below.

Model	Type	Parameters	Choice Acc	Noul F1	Score MAE ↓	Latency P50 ↓	Overall*
openai/gpt-6-luna-decisions	API	Undisclosed	0.922	1.000	0.479	478.4 ms	92.1
nace-ai/drex-v1.5	API	8.95B	0.922	0.989	0.510	598.8 ms	91.4
microsoft/microsoft-decision-1	API	≈9B	0.932	1.000	0.615	560.9 ms	90.9
cloudflare/clef	API	27.36B	0.958	0.979	0.729	559.4 ms	89.8
typesafe/jev-1.13	API	Undisclosed	0.943	0.989	0.760	451.6 ms	89.3
cloudflare/clef-flash	API	9.41B	0.885	1.000	0.677	856.1 ms	88.7
inception/mercury-decide:free	API	Undisclosed	0.964	0.947	0.771	590.4 ms	88.5
xmanii/Mizan-v2-4B-GGUF (Q4_K_M)†	Local	4.54B	0.880	0.990	0.677	214.7 ms	88.1
liquid/d1	API	Undisclosed	0.927	0.946	0.823	437.8 ms	86.6
xmanii/Mizan-v2-4B (BF16)†	Local	4.54B	0.875	0.969	0.740	211.1 ms	86.6
amyrmahdy/decima-base	Local	≈322M	0.479	0.716	0.958	2539.2 ms	62.5
manjunathshiva/opendecider-nano	Local	394.8M	0.396	0.652	0.969	4936.4 ms	57.5
amyrmahdy/decima-small	Local	≈122M	0.469	0.550	0.917	606.3 ms	57.1
convaiinnovations/laya-multilingual	Local	321.9M	0.448	0.603	1.031	901.1 ms	56.9
perplexity/pplx-decider-v1.1-27b	API	26.09B	0.417	0.588	1.500	544.0 ms	50.2
SupersonicLabs/Julia-1	Local	144.3M	0.380	0.436	1.062	367.7 ms	48.7
Dibachain/DibaOne-M3	Local	278M	0.406	0.472	1.260	761.2 ms	48.6
mrm8488/embeddinggemma-2-ft-typed-decision	Local	272.1M	0.391	0.333	0.969	4290.4 ms	46.7
Dibachain/DibaOne-X1	Local	≈395.5M	0.474	0.437†	—	626.8 ms	—
IJyad/jeb-typed-decisions	Local	177.6M	0.401	0.667†	—	508.3 ms	—
Phocinae/Phocinae-Largha-150M-v1	Local	144.3M	0.453	0.449	N/A — incompatible rubric	284.4 ms	—
respan/span-01-lite	API	Undisclosed	—	0.980	—	710.9 ms	—

† Mizan v2 rows are community-reported and not independently verified; the same marker on a Noul metric identifies a binary Choice adapter. Coverage, prediction files, parameter evidence, and submission provenance are in Detailed Results. Candidates without any valid measured primitive remain status-only entries in the detailed model registry; they are not assigned zero scores.

Key Findings

Among benchmark-run results, the highest-scoring complete hosted model is openai/gpt-6-luna-decisions; the best complete local model is amyrmahdy/decima-base. The best complete local checkpoint below 200M parameters is amyrmahdy/decima-small. The Mizan v2 rows are community-reported, not independently verified. These results use only 96 synthetic cases and do not establish real-world decision quality.

🏆 Benchmark Your Model & Submit Results

Have a Persian decision model? You can contribute results without reproducing our full benchmarking infrastructure. We welcome local checkpoints, quantized models, and hosted APIs, including models that support only some decision types.

1. Get the official cases. Load the cases configuration and screening split from this dataset. The copy-paste download command and schema are in the submission guide.
2. Run your model on the exact 96 screening cases. Keep each shared state and question text, Choice option IDs/descriptions, and ordered Score anchors unchanged. Never send gold labels to the model.
3. Save one row per case/question in the same flattened schema as predictions_screening.parquet. The downloadable template shows every field. For unsupported questions, use a null prediction/distribution and mark them unsupported; report those metrics as N/A.
4. Calculate metrics with the benchmark's calculate_metrics function. A working Python example and the exact Choice, Noul, Score, and Overall definitions are in the guide.
5. Measure latency consistently. Report P50/P95 milliseconds per case, excluding model load/download but including input preparation and complete output parsing. The guide covers warm-up, CUDA synchronization, batching, and hosted API network time.
6. Open a submission pull request with the prediction file and the model details listed below. Mizan v2's community submission is an example.

Include the model ID, parameter count and license; variant, precision and quantization; Choice Accuracy, Noul F1, Score MAE, Overall, and valid-decision count; P50/P95 latency; hardware, runtime, batch size and timing method; and whether the model is zero-shot or was trained/tuned on Tasmim-Bench data. Mark unsupported metrics N/A rather than zero. Overall is comparable only when all three native primitives have valid results under the official protocol.

Accepted submissions are labeled community-reported unless we independently reproduce them. State whether the run predates public benchmark access or may have been exposed to public cases; post-publication or benchmark-tuned results are not comparable to pre-publication zero-shot results. We review provenance and schema, but contributors do not need our full model-loading setup.

Full guide, evaluator example, latency methodology, FAQs, and PR template →

Benchmark

• 120 cases: 24 adapter-development and 96 screening cases, with 384 screening decisions and 480 decisions overall.
• Eight workflows use synthetic scenarios and executable oracle rules under fictional policies. Challenge variants include colloquial Persian, Finglish, code-switching, digits, typos, negation, taarof, sarcasm, and distractors.
• There are 0 human-verified labels and 0 independently evaluation-eligible records. Results are a small synthetic pilot, not evidence of operational performance.
• Public screening labels can create future benchmark contamination. Do not use these scores to authorize autonomous medical, legal, financial, employment, security, or government decisions.

The dataset preserves the typed-decision schema: Choice options have stable IDs and descriptions, Score anchors are ordered, and Noul questions state the positive proposition explicitly. The project-owned synthetic cases, fictional policies, and benchmark metadata are MIT-licensed. Third-party model weights and their terms remain governed by their own repositories; the project license does not grant rights to third-party material.

Detailed Results & Downloads

All files and prior releases are preserved. The current versioned snapshot and detailed analyses are available here:

• Screening cases · All screening predictions · Hosted API predictions
• Full-native leaderboard CSV · All candidate metrics · Per-primitive metrics · Hosted API metrics
• Complete model registry and status · Persian robustness · Option robustness · Efficiency
• Evaluation protocol · Reproduction guide · Benchmark manifest · Checksums · Verified parameter metadata and evidence
• Latency addendum v1.0.6: per-model P50/P95 summary · raw per-case samples · API latency cost ledger · methodology · manifest · checksums
• Community Mizan v2 predictions: Parquet · CSV · reported metrics and coverage · submission provenance

Reproduce / Citation

Rebuild and validate this release from saved predictions with the project CLI:

uv run python -m tasmim_v1.bench.publish --repo mshojaei77/tasmim-bench --version v1.0.4 --update-leaderboard --validate

@misc{tasmim_bench_v1,
  title = {Tasmim-Bench: Persian Typed-Decision Model Benchmark},
  author = {Tasmim-v1 contributors},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mshojaei77/tasmim-bench}},
  note = {Version v1.0.4}
}