Tasmim-Bench

Persian Typed-Decision Model Benchmark

Evaluating small Decision Models on Persian Choice, Score, and Noul tasks.

Overview

Tasmim-Bench is a small, reproducible zero-shot screen of structured decision models on Persian-language cases. It evaluates decisions conditioned on a shared state and typed questions with dynamic options or ordered score anchors. It is a dataset and results repository; it does not host an application or Hugging Face Space.

Benchmark Design

The v1 pilot has 120 cases, 480 labeled typed decisions, and 40 semantic families. The saved, family-separated splits are 24 adapter-development cases and 96 screening cases. Each case has two Choice questions, one Score question, and one Noul proposition. The screening set therefore has 192 Choice, 96 Noul, and 96 Score decisions. Cases are synthetic; oracle labels are executable deterministic rules under fictional policies. There are 0 human-verified labels and 0 independently evaluation-eligible records.

Eight workflows are represented: customer support, agent tool selection, RAG evidence, content moderation, e-commerce returns, document review, invoice processing, and security triage. Persian challenge variation includes formal and colloquial Persian, Finglish, Persian-English code-switching, digits, half-spaces, typos, indirect requests/taarof, sarcasm, negation, distractors, long context, underspecification, competing intents, and dynamic options.

The typed schema keeps state, questions, and gold as JSON-encoded strings in Parquet. Choice criteria are ID-to-description mappings, Score criteria are ordered anchors, and Noul asks an explicitly worded true/false proposition. See schema.json, prediction_schema.json, and oracle_evidence.jsonl.

Main Leaderboard

Scores below use the 96-case screening set only. Composite rank is limited to full, native Choice+Score+Noul models with complete valid prediction coverage. The normalized pilot composite is 100 × mean(Choice accuracy, Noul true-class F1, 1 − Score MAE / 3). It is a transparent sorting aid, not a validated utility function. Ranking differences are uncertain on this small synthetic sample; read the per-primitive results and intervals.

Rank	Model	Parameters	Choice Acc	Noul F1	Score MAE	Overall*	Coverage	P95 latency (s)	Peak VRAM	License
1	amyrmahdy/decima-base	≈322.0M	0.479	0.716	0.958	62.517	384/384	N/A	N/A	apache-2.0
2	manjunathshiva/opendecider-nano	394.8M	0.396	0.652	0.969	57.481	384/384	N/A	N/A	apache-2.0
3	amyrmahdy/decima-small	≈122.0M	0.469	0.550	0.917	57.106	384/384	N/A	N/A	apache-2.0
4	convaiinnovations/laya-multilingual	321.9M	0.448	0.603	1.031	56.911	384/384	N/A	N/A	apache-2.0
5	SupersonicLabs/Julia-1	144.3M	0.380	0.436	1.062	48.731	384/384	N/A	N/A	apache-2.0
6	Dibachain/DibaOne-M3	278.0M	0.406	0.472	1.260	48.601	384/384	N/A	N/A	cc-by-nc-sa-4.0
7	mrm8488/embeddinggemma-2-ft-typed-decision	272.1M	0.391	0.333	0.969	46.701	384/384	N/A	N/A	apache-2.0

The full table for all 19 candidates, including unavailable and partially supported entries, is leaderboard.csv. ≈ marks parameter counts estimated from pinned model cards or combined components. Peak VRAM is N/A because inference ran on CPU with a CPU-only PyTorch build.

Specialized Leaderboards

Choice accuracy

Rank	Model	Choice accuracy	Decisions	Support	License
1	amyrmahdy/decima-base	0.479	192	native	apache-2.0
2	Dibachain/DibaOne-X1	0.474	192	native	apache-2.0
3	amyrmahdy/decima-small	0.469	192	native	apache-2.0
4	Phocinae/Phocinae-Largha-150M-v1	0.453	192	native	apache-2.0
5	convaiinnovations/laya-multilingual	0.448	192	native	apache-2.0
6	Dibachain/DibaOne-M3	0.406	192	native	cc-by-nc-sa-4.0
7	IJyad/jeb-typed-decisions	0.401	192	native	apache-2.0
8	manjunathshiva/opendecider-nano	0.396	192	native	apache-2.0
9	mrm8488/embeddinggemma-2-ft-typed-decision	0.391	192	native	apache-2.0
10	SupersonicLabs/Julia-1	0.380	192	native	apache-2.0

Noul positive-class F1

Rank	Model	Noul true F1	Decisions	Support	License
1	amyrmahdy/decima-base	0.716	96	native	apache-2.0
2	IJyad/jeb-typed-decisions	0.667	96	adapted	apache-2.0
3	manjunathshiva/opendecider-nano	0.652	96	native	apache-2.0
4	convaiinnovations/laya-multilingual	0.603	96	native	apache-2.0
5	amyrmahdy/decima-small	0.550	96	native	apache-2.0
6	Dibachain/DibaOne-M3	0.472	96	native	cc-by-nc-sa-4.0
7	Phocinae/Phocinae-Largha-150M-v1	0.449	96	native	apache-2.0
8	Dibachain/DibaOne-X1	0.437	96	adapted	apache-2.0
9	SupersonicLabs/Julia-1	0.436	96	native	apache-2.0
10	mrm8488/embeddinggemma-2-ft-typed-decision	0.333	96	native	apache-2.0

Score ordinal error

Rank	Model	Score MAE	Decisions	Support	License
1	amyrmahdy/decima-small	0.917	96	native	apache-2.0
2	amyrmahdy/decima-base	0.958	96	native	apache-2.0
3	mrm8488/embeddinggemma-2-ft-typed-decision	0.969	96	native	apache-2.0
4	manjunathshiva/opendecider-nano	0.969	96	native	apache-2.0
5	convaiinnovations/laya-multilingual	1.031	96	native	apache-2.0
6	SupersonicLabs/Julia-1	1.062	96	native	apache-2.0
7	Dibachain/DibaOne-M3	1.260	96	native	cc-by-nc-sa-4.0

Only measured predictions with valid distributions and labels appear in primitive rankings. Adapted Noul results are marked as adapted by the capability field in the CSV; they are not native full-typed entries.

Persian challenge slices

These slice counts are small. The numbers are descriptive, not stable rankings.

Rank by Choice	Model	Screening slice	Choice accuracy	Noul true F1	Score MAE
1	manjunathshiva/opendecider-nano	colloquial_persian	0.583 (n=12)	0.800 (n=6)	1.000 (n=6)
2	amyrmahdy/decima-base	colloquial_persian	0.500 (n=12)	0.750 (n=6)	0.667 (n=6)
3	Phocinae/Phocinae-Largha-150M-v1	colloquial_persian	0.417 (n=12)	0.500 (n=6)	N/A
4	amyrmahdy/decima-small	colloquial_persian	0.417 (n=12)	0.750 (n=6)	0.833 (n=6)
5	Dibachain/DibaOne-M3	colloquial_persian	0.417 (n=12)	0.500 (n=6)	2.167 (n=6)
6	convaiinnovations/laya-multilingual	colloquial_persian	0.333 (n=12)	0.500 (n=6)	1.333 (n=6)
7	IJyad/jeb-typed-decisions	colloquial_persian	0.333 (n=12)	0.909 (n=6)	N/A
8	mrm8488/embeddinggemma-2-ft-typed-decision	colloquial_persian	0.250 (n=12)	0.286 (n=6)	0.667 (n=6)
9	SupersonicLabs/Julia-1	colloquial_persian	0.250 (n=12)	0.571 (n=6)	1.333 (n=6)
10	Dibachain/DibaOne-X1	colloquial_persian	0.250 (n=12)	0.500 (n=6)	N/A
1	amyrmahdy/decima-small	finglish	0.750 (n=12)	0.000 (n=6)	0.667 (n=6)
2	amyrmahdy/decima-base	finglish	0.667 (n=12)	0.667 (n=6)	0.833 (n=6)
3	Dibachain/DibaOne-X1	finglish	0.667 (n=12)	0.000 (n=6)	N/A
4	convaiinnovations/laya-multilingual	finglish	0.500 (n=12)	0.500 (n=6)	1.167 (n=6)
5	Phocinae/Phocinae-Largha-150M-v1	finglish	0.500 (n=12)	0.333 (n=6)	N/A
6	mrm8488/embeddinggemma-2-ft-typed-decision	finglish	0.500 (n=12)	0.667 (n=6)	0.667 (n=6)
7	Dibachain/DibaOne-M3	finglish	0.500 (n=12)	0.667 (n=6)	0.667 (n=6)
8	manjunathshiva/opendecider-nano	finglish	0.500 (n=12)	0.571 (n=6)	0.833 (n=6)
9	SupersonicLabs/Julia-1	finglish	0.417 (n=12)	0.000 (n=6)	1.167 (n=6)
10	IJyad/jeb-typed-decisions	finglish	0.417 (n=12)	0.500 (n=6)	N/A
1	amyrmahdy/decima-small	mixed_persian_english	0.571 (n=14)	0.750 (n=7)	1.143 (n=7)
2	Dibachain/DibaOne-M3	mixed_persian_english	0.500 (n=14)	0.400 (n=7)	1.000 (n=7)
3	Phocinae/Phocinae-Largha-150M-v1	mixed_persian_english	0.429 (n=14)	0.571 (n=7)	N/A
4	mrm8488/embeddinggemma-2-ft-typed-decision	mixed_persian_english	0.429 (n=14)	0.333 (n=7)	1.000 (n=7)
5	convaiinnovations/laya-multilingual	mixed_persian_english	0.357 (n=14)	0.500 (n=7)	0.857 (n=7)
6	amyrmahdy/decima-base	mixed_persian_english	0.357 (n=14)	0.571 (n=7)	1.143 (n=7)
7	Dibachain/DibaOne-X1	mixed_persian_english	0.357 (n=14)	0.750 (n=7)	N/A
8	SupersonicLabs/Julia-1	mixed_persian_english	0.286 (n=14)	0.400 (n=7)	1.286 (n=7)
9	manjunathshiva/opendecider-nano	mixed_persian_english	0.286 (n=14)	0.727 (n=7)	1.000 (n=7)
10	IJyad/jeb-typed-decisions	mixed_persian_english	0.286 (n=14)	0.727 (n=7)	N/A

All saved slice metrics (including formal Persian, typos, negation, sarcasm, distractors, and long context) are in robustness_results.csv.

Option permutation and ID-renaming robustness

| Rank | Model | Perturbation | Decisions | Same argmax | Mean |probability change| | |---|---|---|---|---|---| | 1 | mrm8488/embeddinggemma-2-ft-typed-decision | choice_id_rename | 64 | 0.828 | 0.051983 | | 2 | amyrmahdy/decima-small | choice_id_rename | 64 | 0.812 | 0.079633 | | 3 | amyrmahdy/decima-base | choice_id_rename | 64 | 0.781 | 0.095757 | | 1 | amyrmahdy/decima-base | option_order | 64 | 1.000 | 0.000000 | | 2 | mrm8488/embeddinggemma-2-ft-typed-decision | option_order | 64 | 1.000 | 0.000000 | | 3 | amyrmahdy/decima-small | option_order | 64 | 1.000 | 0.000000 |

These probes use 32 screening semantic families and 64 Choice decisions per variant for each model shown. They compare outputs after remapping renamed option IDs.

Lightweight candidates below 200M parameters

Rank	Model	Parameters	Status	Choice accuracy	Noul F1	Score MAE
1	amyrmahdy/decima-small	≈122.0M	EVALUATED	0.469	0.550	0.917
2	SupersonicLabs/Julia-1	144.3M	EVALUATED	0.380	0.436	1.062
N/A	Phocinae/Phocinae-Largha-150M-v1	144.3M	PARTIAL_CAPABILITY	0.453	0.449	N/A
N/A	ParallaxOpen/Vela-Decision-170M	170.6M	PARTIAL_CAPABILITY	N/A	N/A	N/A
N/A	heman10x/rlcd-modernbert-151m	151.4M	RESOURCE_BLOCKED	N/A	N/A	N/A
N/A	IJyad/jeb-typed-decisions	177.6M	PARTIAL_CAPABILITY	0.401	0.667	N/A
N/A	hotchpotch/bekko-system-one-v0-68m	≈68.0M	LICENSE_RESTRICTED	N/A	N/A	N/A

Parameter counts marked as estimates are not exact. See model_status.csv.

Declared permissive-license candidates

Apache-2.0 is the license declared in the pinned model metadata; this is not an independent audit of upstream training data or a legal deployment clearance. The separate candidate registry retains all license states.

Rank	Model	Status	License	Composite if full native
1	amyrmahdy/decima-base	EVALUATED	apache-2.0	62.517
2	manjunathshiva/opendecider-nano	EVALUATED	apache-2.0	57.481
3	amyrmahdy/decima-small	EVALUATED	apache-2.0	57.106
4	convaiinnovations/laya-multilingual	EVALUATED	apache-2.0	56.911
5	SupersonicLabs/Julia-1	EVALUATED	apache-2.0	48.731
6	mrm8488/embeddinggemma-2-ft-typed-decision	EVALUATED	apache-2.0	46.701
N/A	Phocinae/Phocinae-Largha-150M-v1	PARTIAL_CAPABILITY	apache-2.0	N/A
N/A	slow-stack/laya-typed-decisions-multilingual	RESOURCE_BLOCKED	apache-2.0	N/A
N/A	fastino/GLiNER2.5-multi-Decide	RESOURCE_BLOCKED	apache-2.0	N/A
N/A	Dibachain/DibaOne-X1	PARTIAL_CAPABILITY	apache-2.0	N/A
N/A	wfzyx/von	RESOURCE_BLOCKED	apache-2.0	N/A
N/A	ParallaxOpen/Vela-Decision-170M	PARTIAL_CAPABILITY	apache-2.0	N/A
N/A	heman10x/rlcd-modernbert-151m	RESOURCE_BLOCKED	apache-2.0	N/A
N/A	com-kotobalabs/open-jev-deberta-v3-large	RESOURCE_BLOCKED	apache-2.0	N/A
N/A	fastino/GLiNER2.5-Decide	RESOURCE_BLOCKED	apache-2.0	N/A
N/A	IJyad/jeb-typed-decisions	PARTIAL_CAPABILITY	apache-2.0	N/A
N/A	jaredpalmer/kev-0.5b	OVER_500M	apache-2.0	N/A

6GB GPU suitability

Not measured. The host had an RTX 3050 Laptop GPU with 6 GiB reported memory, but the installed PyTorch wheel was CPU-only and CUDA was unavailable. No model is ranked as 6GB-GPU suitable from this run.

Evaluation Protocol

All models were evaluated zero-shot at pinned revisions. The adapter-development split was used for interface/adapter checks, not weight, prompt, or calibration tuning. The screening split was not used to tune models or calibrate probabilities. Models ran sequentially on CPU. Native typed outputs were kept distinct from adapted binary interfaces; unsupported Score or Noul primitives were not converted into synthetic scores. Each case latency covers the shared state and four questions; per-question latency was not recorded. Timed failures are not presented as successful inference speed.

Accuracy and F1 use labels, Score MAE uses ordinal distance, and probability metrics are reported only for complete vectors. ECE uses ten bins and is unstable at this sample size. Family bootstrap intervals use 1,000 draws and deterministic seeds. Full details and metric definitions are in evaluation_protocol.md, primitive_results.csv, and efficiency_results.csv.

Reproduce the Results

The existing project CLI and publication command are recorded in reproduce/README.md. The publication CLI rebuilds from saved predictions; it does not run inference or fine-tuning.

from datasets import load_dataset
import json

cases = load_dataset("mshojaei77/tasmim-bench", "cases", split="screening")
row = cases[0]
state = json.loads(row["state"])
questions = json.loads(row["questions"])
gold = json.loads(row["gold"])

Detailed Results

• Cases: development.parquet, screening.parquet, cases.jsonl.
• Predictions: predictions_screening.parquet, predictions_development.parquet, and diagnostic predictions_smoke.parquet.
• Scores: leaderboard.json, primitive_results.csv, results_by_task.csv, results_by_language_variant.csv, model_status.csv, robustness_results.csv, option_robustness.csv, efficiency_results.csv.
• Integrity and provenance: benchmark_manifest.json, model_revisions.json, label_provenance.json, checksums.sha256.

Models Evaluated

All 19 candidates from the screening registry are listed below. Not evaluated and partial statuses are not zero scores.

Priority	Model	Evaluation status	Parameter count audit	Checkpoint weight bytes	Architecture	Languages declared	License	Native primitives confirmed	Adapted	Valid screen decisions
1	convaiinnovations/laya-multilingual	EVALUATED	321908998	643835514	LayaTypedDecisions	multilingual, en, de, fr, es, pt, it, nl, sv, da, nb, ru, pl, tr, ar, he, fa, ur, hi, bn, ta, te, kn, ml, th, vi, id, ms, tl, ja, ko, zh, el, hu, fi, ro, sq, sl, sw, af, cy, am, hy, ka, km, my, mn, lv, is, az, jv	apache-2.0	choice, noul, score	None	384
2	amyrmahdy/decima-base	EVALUATED	Model card claims about 322M; exact tensor count unavailable in pinned ONNX metadata.	2906521287	Decima	en, fa, ar, ru, de, fr, es, pt, tr, hi, zh, ja, ko, ur, vi, th	apache-2.0	choice, noul, score	None	384
3	Phocinae/Phocinae-Largha-150M-v1	PARTIAL_CAPABILITY	144292870	288603268	ModernBertForMaskedLM	en, zh	apache-2.0	choice, noul	None	288
4	mrm8488/embeddinggemma-2-ft-typed-decision	EVALUATED	272120348	1088537684	GemmaDecisionModel	en, multilingual	apache-2.0	choice, noul, score	None	384
5	amyrmahdy/decima-small	EVALUATED	Model card claims about 122M; exact tensor count unavailable in pinned ONNX metadata.	1124841990	Decima	en, fa, ar, ru, de, fr, es, pt, tr, hi, ta, zh, ja, ko, sw, ur, vi, th, el, bg	apache-2.0	choice, noul, score	None	384
6	Dibachain/DibaOne-M3	LICENSE_RESTRICTED	278043648	3142419079	XLMRobertaModel	fa, en	cc-by-nc-sa-4.0	choice, noul, score	None	384
7	SupersonicLabs/Julia-1	EVALUATED	144292870	577189056	Not available in pinned config	Not declared in pinned card metadata	apache-2.0	choice, noul, score	None	384
8	slow-stack/laya-typed-decisions-multilingual	RESOURCE_BLOCKED	421293830	1685218440	Not available in pinned config	Not declared in pinned card metadata	apache-2.0	None verified	None	0
9	fastino/GLiNER2.5-multi-Decide	RESOURCE_BLOCKED	287355159	1149461028	Gliner2ForSchemaExtraction	multilingual, en	apache-2.0	None verified	None	0
10	Dibachain/DibaOne-X1	PARTIAL_CAPABILITY	278,044,417 main + about 117.5M stage1/retrieval = about 395.5M active.	4462066912	XLMRobertaModel	fa, en	apache-2.0	choice	choice, noul	288
11	wfzyx/von	RESOURCE_BLOCKED	394781696	3160460147	ModernBertModel	en	apache-2.0	None verified	None	0
12	manjunathshiva/opendecider-nano	EVALUATED	394781696	791686130	ModernBertModel	en	apache-2.0	choice, noul, score	None	384
13	ParallaxOpen/Vela-Decision-170M	PARTIAL_CAPABILITY	170569738	684404049	VelaDecisionForOptionScoring	Not declared in pinned card metadata	apache-2.0	choice	noul	0
14	heman10x/rlcd-modernbert-151m	RESOURCE_BLOCKED	151378176	1515637568	GLiClassModel	Not declared in pinned card metadata	apache-2.0	None verified	None	0
15	com-kotobalabs/open-jev-deberta-v3-large	RESOURCE_BLOCKED	434,012,160 backbone + about 3.15M decision head = about 437.2M active.	1748685796	DebertaV2Model	en	apache-2.0	None verified	None	0
16	fastino/GLiNER2.5-Decide	RESOURCE_BLOCKED	486444053	1945828140	Gliner2ForSchemaExtraction	en	apache-2.0	None verified	None	0
17	IJyad/jeb-typed-decisions	PARTIAL_CAPABILITY	177609985	710466028	ArabicDecisionModel	ar	apache-2.0	choice	choice, noul	288
18	hotchpotch/bekko-system-one-v0-68m	LICENSE_RESTRICTED	Model card reports about 68M; exact tensor count unavailable.	468941361	Not available in pinned config	en	not declared	None verified	None	0
19	jaredpalmer/kev-0.5b	OVER_500M	502,830,976 combined base plus LoRA parameters; over the requested cap.	37074383	Not available in pinned config	en	apache-2.0	None verified	None	0

The pinned revision for every candidate is in model_revisions.json. Resource-blocked, incompatible, license-restricted, over-limit, and partial-capability candidates remain explicitly visible in the results.

Limitations and Responsible Use

• This is a small pilot benchmark, not a definitive model evaluation.
• Synthetic fictional-policy oracle labels do not establish real-world accuracy.
• No labels are human-verified; no records are independently evaluation-eligible.
• The benchmark is Persian-focused but remains small and may not represent Persian dialects, writing contexts, or users.
• Healthcare, legal, financial, hiring, government, security, or other consequential decisions are not validated for autonomous use.
• Small-sample rankings and calibration estimates are uncertain. Many challenge slices contain few semantic families.
• Public screening answers can contaminate future model training and benchmark tuning.
• Models have unequal native primitive coverage; partial results are not full typed-decision comparisons.
• Runtime is CPU-only and VRAM was not measured, so 6GB GPU suitability is unknown.
• Model license fields reflect pinned Hub metadata only; upstream training-data rights and commercial deployment requirements need separate review.

License and Citation

Project-owned synthetic cases, fictional policies, oracle rules, and benchmark metadata are licensed under MIT as stated in the included project LICENSE. Third-party model weights and model-related terms remain governed by each linked model repository. The MIT license does not change those terms.

@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 v1.0.1}
}

Version History

Version	Cases	Screening	Notes
v1.0.1	120	96	Initial public pilot; SHA-256 manifest included.

Native Hugging Face Benchmark Status

No eval.yaml is published for this version. The Hub’s maintained framework identifiers do not include Tasmim’s custom evaluator, and representing the run as another framework would be inaccurate. Native benchmark registration is therefore unavailable for this evaluator; the complete Dataset Card leaderboard and machine-readable score tables are published here instead.