Measured on one RTX 4090
Collection
Open models we measured on German tasks on one RTX 4090: retrieval, speech to text, invoice fields. Notes give our numbers. • 23 items • Updated
task_id stringlengths 11 22 | kind stringclasses 5
values | task stringclasses 12
values | lang stringclasses 2
values | metric stringclasses 5
values | family stringclasses 4
values | family_label stringclasses 4
values | state stringclasses 4
values | best_arm stringclasses 14
values | arms_compared int64 1 9 ⌀ | value float64 4.4 100 ⌀ | baseline float64 0 94.1 ⌀ | baseline_name stringclasses 8
values | skill float64 -0.86 1 ⌀ | n int64 116 8.2k ⌀ | answered int64 600 8.2k ⌀ | top_family stringclasses 3
values | mcnemar_vs_top_p stringlengths 1 6 ⌀ | mcnemar_vs_jev_p stringlengths 1 6 ⌀ | p50_ms_per_decision int64 13 240 ⌀ | measured timestamp[s]date 2026-10-01 00:00:00 2026-10-09 00:00:00 ⌀ | model_version stringclasses 14
values | reason stringclasses 6
values | report_url stringclasses 7
values | extra_json stringlengths 2 526 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
route1-scenario-de | route | route: MASSIVE scenario (18 classes) | de | accuracy | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 72.8 | 13.5 | most frequent answer | 0.686 | 2,974 | 2,974 | jev | null | null | 232 | 2026-10-09T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group", "macro_f1": 73.3, "median_input_tokens_per_decision": 510, "tied_with": [], "usd_per_1000_decisions": 0.0214} |
route1-scenario-de | route | route: MASSIVE scenario (18 classes) | de | accuracy | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide | 2 | 48.5 | 13.5 | most frequent answer | 0.404 | 2,974 | 2,974 | jev | <0.001 | <0.001 | 14 | 2026-10-09T00:00:00 | fastino/GLiNER2.5-multi-Decide (revision e173bca1f0c4) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"english_model_extra": {"arm": "gliner2.5-decide", "note": "English-only model on German rows, labelled extra in the report, not eligible as best arm", "skill": 0.41, "tied_with": [], "value": 49}, "macro_f1": 46.4, "tied_with": [], "vs_top_only_arm_right": 188, "vs_top_only_top_right": 912, "vs_top_paired": 2974} |
route1-scenario-de | route | route: MASSIVE scenario (18 classes) | de | accuracy | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 71.8 | 13.5 | most frequent answer | 0.674 | 2,974 | 2,974 | jev | 0.11 | 0.11 | 89 | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 72.3, "tied_with": [], "vs_top_only_arm_right": 161, "vs_top_only_top_right": 192, "vs_top_paired": 2974} |
route1-scenario-de | route | route: MASSIVE scenario (18 classes) | de | accuracy | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | measured | classic-e5-lr | 2 | 86.9 | 13.5 | most frequent answer | 0.848 | 2,974 | 2,974 | jev | <0.001 | <0.001 | 25 | 2026-10-09T00:00:00 | intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 86.1, "tied_with": [], "vs_top_only_arm_right": 559, "vs_top_only_top_right": 141, "vs_top_paired": 2974} |
route1-scenario-en | route | route: MASSIVE scenario (18 classes) | en | accuracy | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 74.1 | 13.5 | most frequent answer | 0.701 | 2,974 | 2,974 | jev | null | null | 231 | 2026-10-09T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group", "macro_f1": 74.2, "median_input_tokens_per_decision": 504, "tied_with": [], "usd_per_1000_decisions": 0.0212} |
route1-scenario-en | route | route: MASSIVE scenario (18 classes) | en | accuracy | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide | 6 | 60.9 | 13.5 | most frequent answer | 0.547 | 2,974 | 2,974 | jev | <0.001 | <0.001 | 26 | 2026-10-09T00:00:00 | fastino/GLiNER2.5-Decide (revision 9d1bfb848cd1) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 60.2, "tied_with": [], "vs_top_only_arm_right": 149, "vs_top_only_top_right": 543, "vs_top_paired": 2974} |
route1-scenario-en | route | route: MASSIVE scenario (18 classes) | en | accuracy | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 70.6 | 13.5 | most frequent answer | 0.661 | 2,974 | 2,974 | jev | <0.001 | <0.001 | 92 | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 71.8, "tied_with": [], "vs_top_only_arm_right": 134, "vs_top_only_top_right": 237, "vs_top_paired": 2974} |
route1-scenario-en | route | route: MASSIVE scenario (18 classes) | en | accuracy | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | measured | classic-e5-lr | 2 | 89.8 | 13.5 | most frequent answer | 0.883 | 2,974 | 2,974 | jev | <0.001 | <0.001 | 23 | 2026-10-09T00:00:00 | intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 89.6, "tied_with": [], "vs_top_only_arm_right": 560, "vs_top_only_top_right": 92, "vs_top_paired": 2974} |
route1-intent2-de | route | route: MASSIVE intent given the gold scenario, 2+ options | de | accuracy | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 90.7 | 50.2 | most frequent answer | 0.812 | 2,622 | 2,622 | jev | null | null | 234 | 2026-10-09T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group", "macro_f1": 86.5, "median_input_tokens_per_decision": 510, "tied_with": [], "usd_per_1000_decisions": 0.0214} |
route1-intent2-de | route | route: MASSIVE intent given the gold scenario, 2+ options | de | accuracy | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide | 2 | 58.9 | 50.2 | most frequent answer | 0.175 | 2,622 | 2,622 | jev | <0.001 | <0.001 | 13 | 2026-10-09T00:00:00 | fastino/GLiNER2.5-multi-Decide (revision e173bca1f0c4) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"english_model_extra": {"arm": "gliner2.5-decide", "note": "English-only model on German rows, labelled extra in the report, not eligible as best arm", "skill": 0.316, "tied_with": [], "value": 65.9}, "macro_f1": 54.9, "tied_with": [], "vs_top_only_arm_right": 67, "vs_top_only_top_right": 899, "vs_top_paired": 2622} |
route1-intent2-de | route | route: MASSIVE intent given the gold scenario, 2+ options | de | accuracy | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 89 | 50.2 | most frequent answer | 0.779 | 2,622 | 2,621 | jev | 0.002 | 0.002 | 84 | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 82.5, "tied_with": [], "vs_top_only_arm_right": 71, "vs_top_only_top_right": 115, "vs_top_paired": 2622} |
route1-intent2-de | route | route: MASSIVE intent given the gold scenario, 2+ options | de | accuracy | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | measured | classic-e5-lr | 2 | 86.9 | 50.2 | most frequent answer | 0.737 | 2,622 | 2,622 | jev | <0.001 | <0.001 | 24 | 2026-10-09T00:00:00 | intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 72.5, "tied_with": [], "vs_top_only_arm_right": 144, "vs_top_only_top_right": 243, "vs_top_paired": 2622} |
route1-intent2-en | route | route: MASSIVE intent given the gold scenario, 2+ options | en | accuracy | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 92 | 50.2 | most frequent answer | 0.838 | 2,622 | 2,622 | jev | null | null | 230 | 2026-10-09T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group", "macro_f1": 88.4, "median_input_tokens_per_decision": 504, "tied_with": [], "usd_per_1000_decisions": 0.0212} |
route1-intent2-en | route | route: MASSIVE intent given the gold scenario, 2+ options | en | accuracy | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide | 6 | 78.3 | 50.2 | most frequent answer | 0.565 | 2,622 | 2,622 | jev | <0.001 | <0.001 | 26 | 2026-10-09T00:00:00 | fastino/GLiNER2.5-Decide (revision 9d1bfb848cd1) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 74.3, "tied_with": [], "vs_top_only_arm_right": 71, "vs_top_only_top_right": 428, "vs_top_paired": 2622} |
route1-intent2-en | route | route: MASSIVE intent given the gold scenario, 2+ options | en | accuracy | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 91.6 | 50.2 | most frequent answer | 0.832 | 2,622 | 2,622 | jev | 0.575 | 0.575 | 88 | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 86.4, "tied_with": [], "vs_top_only_arm_right": 74, "vs_top_only_top_right": 82, "vs_top_paired": 2622} |
route1-intent2-en | route | route: MASSIVE intent given the gold scenario, 2+ options | en | accuracy | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | measured | classic-e5-lr | 2 | 91.2 | 50.2 | most frequent answer | 0.824 | 2,622 | 2,622 | jev | 0.298 | 0.298 | 23 | 2026-10-09T00:00:00 | intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 79.6, "tied_with": [], "vs_top_only_arm_right": 140, "vs_top_only_top_right": 159, "vs_top_paired": 2622} |
route1-fastdec26-en | route | route: fast-decisions dev, single-label tasks with 26 labels or fewer | en | accuracy | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 63 | 31.2 | most frequent answer | 0.461 | 2,500 | 2,500 | gliner | 0.168 | null | 240 | 2026-10-09T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group", "macro_f1": 55.9, "median_input_tokens_per_decision": 544, "tied_with": [], "usd_per_1000_decisions": 0.0228, "vs_top_only_arm_right... |
route1-fastdec26-en | route | route: fast-decisions dev, single-label tasks with 26 labels or fewer | en | accuracy | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide-1b-intext | 6 | 64.4 | 31.2 | most frequent answer | 0.482 | 2,500 | 2,500 | gliner | null | 0.168 | 26 | 2026-10-09T00:00:00 | fastino/GLiNER2.5-Decide-1B (revision 67af949e17be) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 60.4, "tied_with": []} |
route1-fastdec26-en | route | route: fast-decisions dev, single-label tasks with 26 labels or fewer | en | accuracy | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 63.6 | 31.2 | most frequent answer | 0.471 | 2,500 | 2,500 | gliner | 0.483 | 0.309 | 120 | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"macro_f1": 55.3, "tied_with": [], "vs_top_only_arm_right": 319, "vs_top_only_top_right": 338, "vs_top_paired": 2500} |
route1-fastdec26-en | route | route: fast-decisions dev, single-label tasks with 26 labels or fewer | en | accuracy | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | fast-decisions publishes only the dev split, no train split, so the classic classifier has nothing to train on (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/score.md | {"plan": "plan/t3.md"} |
ex1-slots-de | extract | slot extraction (MASSIVE slots) | de | micro F1, threshold 0.5 for GLiNER | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | Jev has no extraction primitive (choice, score, noul only), docs.typesafe.ai/introduction | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/method.md | {} |
ex1-slots-de | extract | slot extraction (MASSIVE slots) | de | micro F1, threshold 0.5 for GLiNER | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi | 1 | 22.5 | 0 | no extraction | 0.225 | 1,000 | null | null | null | null | null | 2026-10-09T00:00:00 | fastino/gliner2.5-multi-v1 (revision cf5593a5d45e) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/score.md | {"exact_frame": 27.5, "exact_frame_baseline": 33.5, "exact_frame_skill": -0.09, "gold_slots": 941, "p50_ms_per_utterance": 19, "precision": 28.8, "recall": 18.5, "skill_note": "baseline F1 is 0, so skill equals F1", "tied_with": []} |
ex1-slots-de | extract | slot extraction (MASSIVE slots) | de | micro F1, threshold 0.5 for GLiNER | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 45.1 | 0 | no extraction | 0.451 | 1,000 | null | null | null | null | null | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/score.md | {"exact_frame": 45.1, "exact_frame_baseline": 33.5, "exact_frame_skill": 0.174, "gold_slots": 941, "p50_ms_per_utterance": 389, "precision": 52.2, "recall": 39.6, "skill_note": "baseline F1 is 0, so skill equals F1", "tied_with": []} |
ex1-slots-de | extract | slot extraction (MASSIVE slots) | de | micro F1, threshold 0.5 for GLiNER | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | open | null | null | null | null | null | null | null | null | null | null | null | null | null | null | a trained slot tagger is possible (MASSIVE train has slot annotations) but is a different model; T3 covers classification only | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/score.md | {"plan": "plan/t3.md"} |
ex1-slots-en | extract | slot extraction (MASSIVE slots) | en | micro F1, threshold 0.5 for GLiNER | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | Jev has no extraction primitive (choice, score, noul only), docs.typesafe.ai/introduction | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/method.md | {} |
ex1-slots-en | extract | slot extraction (MASSIVE slots) | en | micro F1, threshold 0.5 for GLiNER | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2-large | 2 | 38.1 | 0 | no extraction | 0.381 | 1,000 | null | null | null | null | null | 2026-10-09T00:00:00 | fastino/gliner2-large-v1 (revision d7aa8a2850d1) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/score.md | {"exact_frame": 22.4, "exact_frame_baseline": 33.5, "exact_frame_skill": -0.167, "gold_slots": 937, "p50_ms_per_utterance": 30, "precision": 31, "recall": 49.3, "skill_note": "baseline F1 is 0, so skill equals F1", "tied_with": []} |
ex1-slots-en | extract | slot extraction (MASSIVE slots) | en | micro F1, threshold 0.5 for GLiNER | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 48.9 | 0 | no extraction | 0.489 | 1,000 | null | null | null | null | null | 2026-10-09T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0) | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/score.md | {"exact_frame": 44.9, "exact_frame_baseline": 33.5, "exact_frame_skill": 0.171, "gold_slots": 937, "p50_ms_per_utterance": 401, "precision": 51.2, "recall": 46.9, "skill_note": "baseline F1 is 0, so skill equals F1", "tied_with": []} |
ex1-slots-en | extract | slot extraction (MASSIVE slots) | en | micro F1, threshold 0.5 for GLiNER | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | open | null | null | null | null | null | null | null | null | null | null | null | null | null | null | a trained slot tagger is possible (MASSIVE train has slot annotations) but is a different model; T3 covers classification only | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/ex1/score.md | {"plan": "plan/t3.md"} |
pv1-is_req-de | match | pv1 is_req | de | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 88.2 | 85.3 | most frequent answer (yes) | 0.2 | 136 | null | jev | null | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": []} |
pv1-is_req-de | match | pv1 is_req | de | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide | 3 | 85.3 | 85.3 | most frequent answer (yes) | 0 | 136 | null | jev | 0.125 | 0.125 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-multi-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "english_model_extra": {"arm": "gliner2.5-decide", "note": "English-only model on German rows, not eligible as best arm", "skill": -0.05, "tied_with": [], "value": 84.6}, "tied_with": [], "vs_jev_only_arm_right": 0, "vs_jev_only_jev... |
pv1-is_req-de | match | pv1 is_req | de | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 4 | 86.8 | 85.3 | most frequent answer (yes) | 0.1 | 136 | null | jev | 0.5 | 0.5 | null | 2026-10-01T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 0, "vs_jev_only_jev_right": 2, "vs_top_only_arm_right": 0, "vs_top_only_top_right": 2, "vs_top_paired": 136} |
pv1-is_req-de | match | pv1 is_req | de | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"plan": "plan/t3.md"} |
pv1-must-de | match | pv1 must | de | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 93.1 | 94 | most frequent answer (yes) | -0.143 | 116 | null | local-llm | 0.008 | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_top_only_arm_right": 0, "vs_top_only_top_right": 8, "vs_top_paired": 116} |
pv1-must-de | match | pv1 must | de | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide | 3 | 94 | 94 | most frequent answer (yes) | 0 | 116 | null | local-llm | 0.016 | 1 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-multi-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "english_model_extra": {"arm": "gliner2.5-decide-1b-focus", "note": "English-only model on German rows, not eligible as best arm", "skill": 0, "tied_with": [], "value": 94}, "tied_with": ["gliner2.5-multi-decide-focus", "gliner2.5-m... |
pv1-must-de | match | pv1 must | de | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | qwen3-32b | 4 | 100 | 94 | most frequent answer (yes) | 1 | 116 | null | local-llm | null | 0.008 | null | 2026-10-01T00:00:00 | qwen3:32b | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 8, "vs_jev_only_jev_right": 0} |
pv1-must-de | match | pv1 must | de | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"plan": "plan/t3.md"} |
pv1-axis-de | match | pv1 axis | de | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 93.1 | 16.4 | most frequent answer (testauto) | 0.918 | 116 | null | jev | null | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": []} |
pv1-axis-de | match | pv1 axis | de | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide-focus | 3 | 32.8 | 16.4 | most frequent answer (testauto) | 0.196 | 116 | null | jev | <0.001 | <0.001 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-multi-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "english_model_extra": {"arm": "gliner2.5-decide-focus", "note": "English-only model on German rows, not eligible as best arm", "skill": 0.474, "tied_with": [], "value": 56}, "tied_with": [], "vs_jev_only_arm_right": 1, "vs_jev_only... |
pv1-axis-de | match | pv1 axis | de | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 4 | 87.1 | 16.4 | most frequent answer (testauto) | 0.845 | 116 | null | jev | 0.065 | 0.065 | null | 2026-10-01T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 2, "vs_jev_only_jev_right": 9, "vs_top_only_arm_right": 2, "vs_top_only_top_right": 9, "vs_top_paired": 116} |
pv1-axis-de | match | pv1 axis | de | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"plan": "plan/t3.md"} |
pv1-evidence-de | match | pv1 evidence | de | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 81 | 41.4 | most frequent answer (none) | 0.676 | 116 | null | local-llm | 0.077 | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_top_only_arm_right": 4, "vs_top_only_top_right": 12, "vs_top_paired": 116} |
pv1-evidence-de | match | pv1 evidence | de | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide | 3 | 19 | 41.4 | most frequent answer (none) | -0.382 | 116 | null | local-llm | <0.001 | <0.001 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-multi-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "english_model_extra": {"arm": "gliner2.5-decide", "note": "English-only model on German rows, not eligible as best arm", "skill": -0.441, "tied_with": ["gliner2.5-decide-focus", "gliner2.5-decide-intext"], "value": 15.5}, "tied_wit... |
pv1-evidence-de | match | pv1 evidence | de | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | qwen3.8-27b | 4 | 87.9 | 41.4 | most frequent answer (none) | 0.794 | 116 | null | local-llm | null | 0.077 | null | 2026-10-01T00:00:00 | qwen3.8:27b | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 12, "vs_jev_only_jev_right": 4} |
pv1-evidence-de | match | pv1 evidence | de | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"plan": "plan/t3.md"} |
pv1-level-de | match | pv1 level | de | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 79.3 | 34.5 | most frequent answer (3) | 0.684 | 116 | null | jev | null | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": []} |
pv1-level-de | match | pv1 level | de | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide-intext | 3 | 24.1 | 34.5 | most frequent answer (3) | -0.158 | 116 | null | jev | <0.001 | <0.001 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-multi-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "english_model_extra": {"arm": "gliner2.5-decide", "note": "English-only model on German rows, not eligible as best arm", "skill": 0.066, "tied_with": ["gliner2.5-decide-focus"], "value": 38.8}, "tied_with": [], "vs_jev_only_arm_rig... |
pv1-level-de | match | pv1 level | de | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 4 | 77.6 | 34.5 | most frequent answer (3) | 0.658 | 116 | null | jev | 0.824 | 0.824 | null | 2026-10-01T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 9, "vs_jev_only_jev_right": 11, "vs_top_only_arm_right": 9, "vs_top_only_top_right": 11, "vs_top_paired": 116} |
pv1-level-de | match | pv1 level | de | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-de-adjudicated.md | {"plan": "plan/t3.md"} |
pv1-is_req-en | match | pv1 is_req | en | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 88.5 | 85.6 | most frequent answer (yes) | 0.2 | 139 | null | local-llm | 1 | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_top_only_arm_right": 0, "vs_top_only_top_right": 1, "vs_top_paired": 139} |
pv1-is_req-en | match | pv1 is_req | en | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide | 9 | 85.6 | 85.6 | most frequent answer (yes) | 0 | 139 | null | local-llm | 0.063 | 0.125 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": ["gliner2.5-multi-decide"], "vs_jev_only_arm_right": 0, "vs_jev_only_jev_right": 4, "vs_top_only_arm_right": 0, "vs_top_only_top_right": 5, "vs_top_paired": 139} |
pv1-is_req-en | match | pv1 is_req | en | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 4 | 89.2 | 85.6 | most frequent answer (yes) | 0.25 | 139 | null | local-llm | null | 1 | null | 2026-10-01T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 1, "vs_jev_only_jev_right": 0} |
pv1-is_req-en | match | pv1 is_req | en | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"plan": "plan/t3.md"} |
pv1-must-en | match | pv1 must | en | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 89.1 | 94.1 | most frequent answer (yes) | -0.857 | 119 | null | local-llm | <0.001 | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_top_only_arm_right": 0, "vs_top_only_top_right": 13, "vs_top_paired": 119} |
pv1-must-en | match | pv1 must | en | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide-1b | 9 | 94.1 | 94.1 | most frequent answer (yes) | 0 | 119 | null | local-llm | 0.016 | 0.263 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-Decide-1B | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": ["gliner2.5-decide-1b-focus", "gliner2.5-multi-decide", "gliner2.5-multi-decide-focus", "gliner2.5-multi-decide-intext"], "vs_jev_only_arm_right": 13, "vs_jev_only_jev_right": 7, "vs_top_only_arm_right": 0, "vs_top_only... |
pv1-must-en | match | pv1 must | en | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | qwen3-32b | 4 | 100 | 94.1 | most frequent answer (yes) | 1 | 119 | null | local-llm | null | <0.001 | null | 2026-10-01T00:00:00 | qwen3:32b | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 13, "vs_jev_only_jev_right": 0} |
pv1-must-en | match | pv1 must | en | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"plan": "plan/t3.md"} |
pv1-axis-en | match | pv1 axis | en | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 91.6 | 16 | most frequent answer (ai) | 0.9 | 119 | null | jev | null | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": []} |
pv1-axis-en | match | pv1 axis | en | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide-1b-focus | 9 | 62.2 | 16 | most frequent answer (ai) | 0.55 | 119 | null | jev | <0.001 | <0.001 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-Decide-1B | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 3, "vs_jev_only_jev_right": 38, "vs_top_only_arm_right": 3, "vs_top_only_top_right": 38, "vs_top_paired": 119} |
pv1-axis-en | match | pv1 axis | en | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 4 | 88.2 | 16 | most frequent answer (ai) | 0.86 | 119 | null | jev | 0.289 | 0.289 | null | 2026-10-01T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 2, "vs_jev_only_jev_right": 6, "vs_top_only_arm_right": 2, "vs_top_only_top_right": 6, "vs_top_paired": 119} |
pv1-axis-en | match | pv1 axis | en | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"plan": "plan/t3.md"} |
pv1-evidence-en | match | pv1 evidence | en | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 79 | 37 | most frequent answer (none) | 0.667 | 119 | null | local-llm | 0.503 | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_top_only_arm_right": 8, "vs_top_only_top_right": 12, "vs_top_paired": 119} |
pv1-evidence-en | match | pv1 evidence | en | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide | 9 | 14.3 | 37 | most frequent answer (none) | -0.36 | 119 | null | local-llm | <0.001 | <0.001 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": ["gliner2.5-decide-focus", "gliner2.5-decide-intext"], "vs_jev_only_arm_right": 1, "vs_jev_only_jev_right": 78, "vs_top_only_arm_right": 4, "vs_top_only_top_right": 85, "vs_top_paired": 119} |
pv1-evidence-en | match | pv1 evidence | en | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | qwen3.8-27b | 4 | 82.4 | 37 | most frequent answer (none) | 0.72 | 119 | null | local-llm | null | 0.503 | null | 2026-10-01T00:00:00 | qwen3.8:27b | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 12, "vs_jev_only_jev_right": 8} |
pv1-evidence-en | match | pv1 evidence | en | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"plan": "plan/t3.md"} |
pv1-level-en | match | pv1 level | en | accuracy against the blind reference | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 74.8 | 34.5 | most frequent answer (3) | 0.615 | 119 | null | jev | null | null | null | 2026-10-01T00:00:00 | jev-1.13.0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": []} |
pv1-level-en | match | pv1 level | en | accuracy against the blind reference | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide | 9 | 38.7 | 34.5 | most frequent answer (3) | 0.064 | 119 | null | jev | <0.001 | <0.001 | null | 2026-10-08T00:00:00 | fastino/GLiNER2.5-Decide | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": ["gliner2.5-decide-focus"], "vs_jev_only_arm_right": 8, "vs_jev_only_jev_right": 51, "vs_top_only_arm_right": 8, "vs_top_only_top_right": 51, "vs_top_paired": 119} |
pv1-level-en | match | pv1 level | en | accuracy against the blind reference | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 4 | 70.6 | 34.5 | most frequent answer (3) | 0.551 | 119 | null | jev | 0.442 | 0.442 | null | 2026-10-01T00:00:00 | hf.co/EldanRing/Winnow-12B:Q8_0 | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"baseline_source": "computed from reference/pv1 (same rule as results/pv1/gliner.md)", "tied_with": [], "vs_jev_only_arm_right": 11, "vs_jev_only_jev_right": 16, "vs_top_only_arm_right": 11, "vs_top_only_top_right": 16, "vs_top_paired": 119} |
pv1-level-en | match | pv1 level | en | accuracy against the blind reference | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data for this task; the only labels are the blind reference of the test rows themselves (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/score-en.md | {"plan": "plan/t3.md"} |
einv1-yesno-de | yesno | yes/no question over the free text of an e-invoice (KoSIT XRechnung test suite, 34 invoices, ten questions) | de | accuracy against a blind reference, pooled over 340 cells and per question | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | planned | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | {"plan": "plan/einv1.md"} |
einv1-yesno-de | yesno | yes/no question over the free text of an e-invoice (KoSIT XRechnung test suite, 34 invoices, ten questions) | de | accuracy against a blind reference, pooled over 340 cells and per question | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | planned | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | {"plan": "plan/einv1.md"} |
einv1-yesno-de | yesno | yes/no question over the free text of an e-invoice (KoSIT XRechnung test suite, 34 invoices, ten questions) | de | accuracy against a blind reference, pooled over 340 cells and per question | local-llm | open LLM through Ollama, local GPU (RTX 4090) | planned | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | {"plan": "plan/einv1.md"} |
einv1-yesno-de | yesno | yes/no question over the free text of an e-invoice (KoSIT XRechnung test suite, 34 invoices, ten questions) | de | accuracy against a blind reference, pooled over 340 cells and per question | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no labelled training data; the 34 invoices are the test set (plan/t3.md) | null | {"plan": "plan/t3.md"} |
route1-automation95-de | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 60.3 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.603 | 5,596 | 5,596 | jev | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 81.2, "accuracy_automated": 95.2, "automation90": 78.2, "automation90_baseline": 0, "brier": 0.2853, "crossfit90": 78.3, "crossfit90_accuracy": 90, "crossfit95": 60.3, "crossfit95_accuracy": 95.2, "ece": 0.0817, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.96, "tied_with":... |
route1-automation95-de | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide | 2 | 4.4 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.044 | 5,596 | 5,596 | jev | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 53.4, "accuracy_automated": 95.1, "automation90": 12.7, "automation90_baseline": 0, "brier": 0.6012, "crossfit90": 12.6, "crossfit90_accuracy": 90.2, "crossfit95": 4.8, "crossfit95_accuracy": 94.8, "ece": 0.0668, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.8968, "tied_wit... |
route1-automation95-de | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 58 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.58 | 5,596 | 5,595 | jev | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 79.8, "accuracy_automated": 95, "automation90": 75.4, "automation90_baseline": 0, "brier": 0.3528, "crossfit90": 75.4, "crossfit90_accuracy": 89.9, "crossfit95": 57.9, "crossfit95_accuracy": 95.1, "ece": 0.1576, "letter_mass_median": 0.9998, "letter_mass_p10": 0.9906, "skill_note": "baseline is 0, so s... |
route1-automation95-de | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | measured | classic-e5-lr | 2 | 76.4 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.764 | 5,596 | 5,596 | null | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 86.9, "accuracy_automated": 95, "automation90": 93.3, "automation90_baseline": 0, "brier": 0.2384, "crossfit90": 93.3, "crossfit90_accuracy": 90, "crossfit95": 76.7, "crossfit95_accuracy": 94.8, "ece": 0.1744, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.52, "tied_with": [... |
route1-automation95-en | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 42.6 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.426 | 8,196 | 8,196 | jev | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 76.3, "accuracy_automated": 95.2, "automation90": 64.7, "automation90_baseline": 0, "brier": 0.3614, "crossfit90": 64.7, "crossfit90_accuracy": 90.1, "crossfit95": 42.6, "crossfit95_accuracy": 95.2, "ece": 0.1297, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 1, "tied_with": ... |
route1-automation95-en | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide-1b | 6 | 10.8 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.108 | 8,196 | 8,196 | jev | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 62.6, "accuracy_automated": 95, "automation90": 23.3, "automation90_baseline": 0, "brier": 0.4964, "crossfit90": 23.2, "crossfit90_accuracy": 90.1, "crossfit95": 9.2, "crossfit95_accuracy": 94.7, "ece": 0.061, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.8815, "tied_with":... |
route1-automation95-en | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | winnow-12b | 1 | 41.3 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.413 | 8,196 | 8,096 | jev | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 74.4, "accuracy_automated": 95.1, "automation90": 58.7, "automation90_baseline": 0, "brier": 0.4355, "crossfit90": 58.7, "crossfit90_accuracy": 90, "crossfit95": 41.9, "crossfit95_accuracy": 94.9, "ece": 0.1972, "letter_mass_median": 0.9997, "letter_mass_p10": 0.9903, "skill_note": "baseline is 0, so s... |
route1-automation95-en | automation | automation: route1 (MASSIVE scenario, intent 2+ options; en also fast-decisions) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | measured | classic-e5-lr | 2 | 87.7 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.877 | 5,596 | 5,596 | null | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/route1/automation.md | {"accuracy_all": 90.5, "accuracy_automated": 95, "automation90": 100, "automation90_baseline": 0, "brier": 0.1882, "crossfit90": 100, "crossfit90_accuracy": 90.5, "crossfit95": 87.8, "crossfit95_accuracy": 95, "ece": 0.1651, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.4581, "tied_with": [... |
pv1-automation95-de | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 53.5 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.535 | 600 | 600 | local-llm | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {"accuracy_all": 87, "accuracy_automated": 95.3, "automation90": 90.8, "automation90_baseline": 19.3, "brier": 0.2215, "crossfit90": 91.3, "crossfit90_accuracy": 90, "crossfit95": 53.4, "crossfit95_accuracy": 95, "ece": 0.073, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.92, "tied_with": [... |
pv1-automation95-de | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-multi-decide-focus | 3 | 28.5 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.285 | 600 | 600 | local-llm | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {"accuracy_all": 51.5, "accuracy_automated": 95.9, "automation90": 38.5, "automation90_baseline": 19.3, "brier": 0.6362, "crossfit90": 37.8, "crossfit90_accuracy": 90.8, "crossfit95": 24, "crossfit95_accuracy": 94.5, "ece": 0.1876, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.5652, "tied_w... |
pv1-automation95-de | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | qwen3.8-27b | 4 | 68.2 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.682 | 600 | 600 | local-llm | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {"accuracy_all": 85.2, "accuracy_automated": 95.1, "automation90": 90.5, "automation90_baseline": 19.3, "brier": 0.2046, "crossfit90": 90.2, "crossfit90_accuracy": 89.6, "crossfit95": 70.1, "crossfit95_accuracy": 94.8, "ece": 0.0319, "letter_mass_median": 0.998, "letter_mass_p10": 0.9963, "skill_note": "baseline is 0, ... |
pv1-automation95-de | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | de | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no classic arm on pv1: the classic classifier needs a train split, pv1 has none (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {} |
pv1-automation95-en | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | jev | TypeSafe Jev, hosted API (api.typesafe.ai) | measured | jev | 1 | 51.9 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.519 | 615 | 615 | local-llm | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {"accuracy_all": 84.7, "accuracy_automated": 95.3, "automation90": 79.3, "automation90_baseline": 19.4, "brier": 0.2283, "crossfit90": 83.9, "crossfit90_accuracy": 89.3, "crossfit95": 58, "crossfit95_accuracy": 93.8, "ece": 0.037, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.92, "tied_with... |
pv1-automation95-en | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | gliner | Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090) | measured | gliner2.5-decide-1b-focus | 9 | 24.6 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.246 | 615 | 615 | local-llm | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {"accuracy_all": 52.8, "accuracy_automated": 95.4, "automation90": 38.9, "automation90_baseline": 19.4, "brier": 0.5757, "crossfit90": 38.8, "crossfit90_accuracy": 89.5, "crossfit95": 25.9, "crossfit95_accuracy": 94.3, "ece": 0.1324, "skill_note": "baseline is 0, so skill equals value / 100", "threshold": 0.6556, "tied... |
pv1-automation95-en | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | local-llm | open LLM through Ollama, local GPU (RTX 4090) | measured | qwen3.8-27b | 4 | 56.4 | 0 | most frequent answer per option set, confidence = its in-sample share | 0.564 | 615 | 615 | local-llm | null | null | null | null | null | null | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {"accuracy_all": 82.3, "accuracy_automated": 95.1, "automation90": 85.7, "automation90_baseline": 19.4, "brier": 0.2413, "crossfit90": 84.3, "crossfit90_accuracy": 89.8, "crossfit95": 59.2, "crossfit95_accuracy": 93.6, "ece": 0.0463, "letter_mass_median": 0.9977, "letter_mass_p10": 0.9949, "skill_note": "baseline is 0,... |
pv1-automation95-en | automation | automation: pv1 line questions (is_req, must, axis, evidence, level) | en | automation rate at 95 % accuracy: largest share of items, ranked by the probability of the arm's own answer, whose accuracy stays >= 95 %; threshold chosen in-sample | classic | sentence embeddings plus a trained logistic regression, local (Mac, MPS) | not_applicable | null | null | null | null | null | null | null | null | null | null | null | null | null | null | no classic arm on pv1: the classic classifier needs a train split, pv1 has none (plan/t3.md) | https://github.com/bks-lab/lab-bench/blob/3584dfc8fddeb353916650f6da1f449cf20e2ccb/results/pv1/automation.md | {} |
lab-bench results
Aggregated results of lab-bench, the test bench BKS-Lab uses to decide which model can run a task on hardware we control. Every number here is measured on one RTX 4090 with 24 GB, or comes from a hosted service where the table says so.
This dataset holds the summary tables only. Cases, requests, references, per-question result rows and the code that produced them live in the repository, at the commit named below. Every row links to its report there.
- Built from commit
3584dfc8fddeb353916650f6da1f449cf20e2ccbof bks-lab/lab-bench, tool map generated 2026-10-12. - Write-ups: Test bench and Decision models on bks-lab.com.
Configs
toolmap. One row per task and tool family. Families are a hosted
decision API (TypeSafe Jev), GLiNER encoders, open LLMs through Ollama and a
classic trained classifier. Tasks are routing (MASSIVE 1.1, fast-decisions),
slot extraction (MASSIVE 1.1) and a CV to job posting match. state says
whether a cell is measured, planned, open or not applicable. skill is
(value - baseline) / (100 - baseline). best_arm is chosen after the fact
on the same data, which flatters families with many arms. The p values are
exact McNemar tests against the top family and against Jev, uncorrected.
Fields that are not columns are kept in extra_json.
capabilities. One row per capability test (C1 German retrieval for RAG, C2 German speech to text, C3 invoice fields from page images, more to come). Each row names the best local model, its licence class, the quality value with a 95 % bootstrap interval, the baseline, a sanity check against a published reference value, speed, peak VRAM and GPU energy.
What is not in here, and why
- No per-question outputs of Jev. They are published in the repository with TypeSafe's permission for reproducing the measurements only and must not be used to train or distil a model or to build a competing product, see LICENSE-JEV-OUTPUTS. A dataset on the Hub is a training source, so they stay out. The aggregated scores of Jev are benchmark results and are included.
- No cases. The routing and extraction cases come from MASSIVE 1.1 (Amazon, CC BY 4.0) and fastino/fast-decisions (Apache-2.0). Use them from their publishers.
How to cite the numbers
Run 1 of each arm, one run per arm unless a report says otherwise. A cell compares tool families on one task and says nothing about another task. If a number matters to you, read its report first: the method, the limits and the open questions are written down there.
Licence
The tables are CC BY 4.0, BKS-Lab contributors. Rows derived from MASSIVE 1.1 also credit Amazon (CC BY 4.0), rows derived from fast-decisions credit Fastino (Apache-2.0).
- Downloads last month
- 126