--- pretty_name: "Tasmim-Bench: Persian Typed-Decision Model Benchmark" license: mit language: - fa - en tags: - benchmark - typed-decisions - persian - text task_categories: - text-classification configs: - config_name: cases data_files: - split: development path: versions/v1.1.2/data/development.parquet - split: screening path: versions/v1.1.2/data/screening.parquet - config_name: predictions data_files: - split: smoke path: versions/v1.1.2/data/predictions_smoke.parquet - split: development path: versions/v1.1.2/data/predictions_development.parquet - split: screening path: versions/v1.1.2/data/predictions_screening.parquet - config_name: hosted_api_predictions data_files: - split: screening path: versions/v1.1.2/data/hosted_api_predictions_screening.parquet - config_name: community_mizan_v2_predictions data_files: - split: screening path: versions/v1.0.5/submissions/xmanii-mizan-v2/predictions_screening_mizan.parquet --- # 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 checkpoint metadata or linked reports. Exact hosted checkpoint counts use safetensors-index totals when checkpoint identity is established; rounded author reports and backbone-only estimates are marked with `≈`. Other hosted sizes remain `Undisclosed` when no defensible count maps to the served model. `—` 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 where Tasmim timing is available; `‡` marks author-reported Mizan medians per text request, with question and sample counts unspecified, rather than a Tasmim-controlled case measurement. 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](https://openrouter.ai/openai/gpt-6-luna-decisions) | API | Undisclosed | 0.922 | 1.000 | 0.479 | 478.4 ms | 92.1 | | [nace-ai/drex-v1.5](https://openrouter.ai/nace-ai/drex-v1.5) | API | 8.95B | 0.922 | 0.989 | 0.510 | 598.8 ms | 91.4 | | [Mizan v3 Q4†](https://huggingface.co/Nimruz/Mizan-v3-3B-GGUF) | Local | 3.20B | 0.917 | 0.970 | 0.458 | 55 ms‡ | 91.1 | | [microsoft/microsoft-decision-1](https://openrouter.ai/microsoft/microsoft-decision-1) | API | ≈9B | 0.932 | 1.000 | 0.615 | 560.9 ms | 90.9 | | [cloudflare/clef](https://openrouter.ai/cloudflare/clef) | API | 27.36B | 0.958 | 0.979 | 0.729 | 559.4 ms | 89.8 | | [typesafe/jev-1.13](https://openrouter.ai/typesafe/jev-1.13) | API | Undisclosed | 0.943 | 0.989 | 0.760 | 451.6 ms | 89.3 | | [Mizan 1.0 Q4†](https://huggingface.co/Nimruz/Mizan-v3-3B/tree/bf8becc1073b5e03445e0a9fe766489af3a2a15b) | Local | ≈4.00B | 0.870 | 0.970 | 0.490 | 80 ms‡ | 89.2 | | [cloudflare/clef-flash](https://openrouter.ai/cloudflare/clef-flash) | API | 9.41B | 0.885 | 1.000 | 0.677 | 856.1 ms | 88.7 | | [inception/mercury-decide:free](https://openrouter.ai/inception/mercury-decide%3Afree) | API | Undisclosed | 0.964 | 0.947 | 0.771 | 590.4 ms | 88.5 | | [xmanii/Mizan-v2-4B-GGUF (Q4_K_M)†](https://huggingface.co/xmanii/Mizan-v2-4B-GGUF) | Local | 4.54B | 0.880 | 0.990 | 0.677 | 214.7 ms | 88.1 | | [liquid/d1](https://openrouter.ai/liquid/d1) | API | Undisclosed | 0.927 | 0.946 | 0.823 | 437.8 ms | 86.6 | | [xmanii/Mizan-v2-4B (BF16)†](https://huggingface.co/xmanii/Mizan-v2-4B) | Local | 4.54B | 0.875 | 0.969 | 0.740 | 211.1 ms | 86.6 | | [amyrmahdy/decima-base](https://huggingface.co/amyrmahdy/decima-base) | Local | ≈322M | 0.479 | 0.716 | 0.958 | 2539.2 ms | 62.5 | | [manjunathshiva/opendecider-nano](https://huggingface.co/manjunathshiva/opendecider-nano) | Local | 394.8M | 0.396 | 0.652 | 0.969 | 4936.4 ms | 57.5 | | [amyrmahdy/decima-small](https://huggingface.co/amyrmahdy/decima-small) | Local | ≈122M | 0.469 | 0.550 | 0.917 | 606.3 ms | 57.1 | | [convaiinnovations/laya-multilingual](https://huggingface.co/convaiinnovations/laya-multilingual) | Local | 321.9M | 0.448 | 0.603 | 1.031 | 901.1 ms | 56.9 | | [Lukitaduarte/dinah-0](https://huggingface.co/Lukitaduarte/dinah-0) | Local | 149.6M | 0.406 | 0.667 | 1.104 | 4478.1 ms | 56.8 | | [hotchpotch/bekko-system-one-v0-17m](https://huggingface.co/hotchpotch/bekko-system-one-v0-17m) | Local | ≈17M | 0.250 | 0.681 | 0.979 | 118.9 ms | 53.5 | | [perplexity/pplx-decider-v1.1-27b](https://openrouter.ai/perplexity/pplx-decider-v1.1-27b) | API | 26.09B | 0.417 | 0.588 | 1.500 | 544.0 ms | 50.2 | | [SupersonicLabs/Julia-1](https://huggingface.co/SupersonicLabs/Julia-1) | Local | 144.3M | 0.380 | 0.436 | 1.062 | 367.7 ms | 48.7 | | [Dibachain/DibaOne-M3](https://huggingface.co/Dibachain/DibaOne-M3) | Local | 278M | 0.406 | 0.472 | 1.260 | 761.2 ms | 48.6 | | [kshitijthakkar/nirnaya-0.4b-0.2a-decision](https://huggingface.co/kshitijthakkar/nirnaya-0.4b-0.2a-decision) | Local | 390M | 0.323 | 0.574 | 1.385 | 9553.0 ms | 47.8 | | [mo-shadfar/laya-fa-general](https://huggingface.co/mo-shadfar/laya-fa-general) | Local | 321.9M | 0.490 | 0.255 | 1.000 | N/A | 47.0 | | [mrm8488/embeddinggemma-2-ft-typed-decision](https://huggingface.co/mrm8488/embeddinggemma-2-ft-typed-decision) | Local | 272.1M | 0.391 | 0.333 | 0.969 | 4290.4 ms | 46.7 | | [Manav2op/verdict-small](https://huggingface.co/Manav2op/verdict-small) | Local | 117.7M | 0.406 | 0.078 | 1.146 | 193.5 ms | 36.8 | | [hotchpotch/bekko-system-one-v0-68m](https://huggingface.co/hotchpotch/bekko-system-one-v0-68m) | Local | ≈68M | 0.365 | 0.000 | 0.875 | 970.8 ms | 35.8 | | [Dibachain/DibaOne-X1](https://huggingface.co/Dibachain/DibaOne-X1) | Local | ≈395.5M | 0.474 | 0.437† | — | 626.8 ms | — | | [IJyad/jeb-typed-decisions](https://huggingface.co/IJyad/jeb-typed-decisions) | Local | 177.6M | 0.401 | 0.667† | — | 508.3 ms | — | | [Phocinae/Phocinae-Largha-150M-v1](https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1) | Local | 144.3M | 0.453 | 0.449 | N/A — incompatible rubric | 284.4 ms | — | | [tasksource/tasksource-decider-nano](https://huggingface.co/tasksource/tasksource-decider-nano) | Local | 149M | 0.417 | 0.575 | — | 1288.8 ms | — | | [respan/span-01-lite](https://openrouter.ai/respan/span-01-lite) | API | Undisclosed | — | 0.980 | — | 710.9 ms | — | **Prior checkpoint training:** - [mo-shadfar/laya-fa-general](https://huggingface.co/mo-shadfar/laya-fa-general): Fine-tuned checkpoint; model card reports about 300 Persian customer-service cases; rl_agent_config.json marks fine_tuned=true. This checkpoint was already fine-tuned before this screening run; its result is not a zero-shot base-model result. † Mizan rows are community-reported and not independently verified; the same marker on a Noul metric identifies a binary Choice adapter. The Nimruz report supplies aggregate metrics for v1.0.4 (96 cases) but no per-question predictions or valid-decision count; its author notes format familiarity and small-sample uncertainty. `‡` marks its model-card median reports: 55 ms for Mizan v3 Q4 and 80 ms for Mizan 1.0 Q4 on an NVIDIA L4; these are per text request, not standardized Tasmim case timings. Coverage, available 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](https://openrouter.ai/openai/gpt-6-luna-decisions); the best complete local model is [amyrmahdy/decima-base](https://huggingface.co/amyrmahdy/decima-base). The best complete local checkpoint below 200M parameters is [amyrmahdy/decima-small](https://huggingface.co/amyrmahdy/decima-small). The model card also reports Mizan v3 Q4 at 91.1 on Tasmim-Bench v1.0.4; that result is community-reported and was not independently reproduced. Community-reported Mizan results have not been independently verified. These results use only 96 synthetic screening 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](docs/SUBMIT_RESULTS.md). 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`](versions/v1.0.4/data/predictions_screening.parquet). The [downloadable template](docs/prediction_submission_template.jsonl) 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](https://huggingface.co/datasets/mshojaei77/tasmim-bench/discussions/new?type=pull_request)** with the prediction file and the model details listed below. [Mizan v2's community submission](https://huggingface.co/datasets/mshojaei77/tasmim-bench/discussions/1) 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 →](docs/SUBMIT_RESULTS.md) ## 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. - Candidate discovery also reviewed all 159 rows in the pinned [Decision Index registry](https://huggingface.co/spaces/multimodalart/jev-decision-index) at revision `8b5eee110f65bc5e0ceb8eb4231fba1a7cd21689`. The expansion added 23 unique sub-1B candidates; language tags alone were not used as an exclusion. Explicit non-English exclusions are listed in the audit. See the complete [eligibility audit](), [registry](), [execution status](), and [capability evidence](). - 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]() - [Decision Index expansion]() · [full registry]() · [eligibility audit]() · [candidate capabilities]() · [per-model/per-primitive metrics]() · [raw timing samples]() - [Evaluation protocol]() · [Reproduction guide]() · [Benchmark manifest]() · [Checksums]() · [Verified parameter metadata and evidence]() · [Additional size and latency evidence]() - Latency addendum v1.0.7: [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]() - Additional community metrics (v1.1.1): [reported results]() · [submission provenance]() · [checksums](). Aggregate values only; no per-question predictions were supplied. ## Reproduce / Citation Rebuild and validate this release from saved predictions with the project CLI: ```bash uv run python -m tasmim_v1.bench.publish --repo mshojaei77/tasmim-bench --version v1.1.2 --update-leaderboard --validate ``` ```bibtex @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 v1.1.2} } ```