Datasets:

mlboydaisuke
/

coreai-decision-models-public231

Tags:

License:

Dataset card Files Files and versions

xet

Community

The Dataset Viewer has been disabled on this dataset.

Decision models on Core AI, 231 public items

Eight models from the coreai-kit catalog answered the 231 public items of JevBench (easy 48, standard 72, hard 111) with Core AI on a Mac: Apple M4 Max, int8 bundles, except the laya-multilingual encoder in fp16. Six are decision models; two are chat models asked zero-shot. JevBench's own harness at commit 2fa63fa (v1.4.0) sent each item through its stock typesafe adapter to decide-cli serve from coreai-kit at commit adbc755, one question per request. JevBench's own code scored the answers. Every per-item result is here, so the accuracy, calibration and latency figures can be recomputed from this dataset.

This is an independent run, not an official JevBench result.

Results

model (kit catalog id, Mac int8, contended)	easy 48	standard 72	hard 111	ECE hard	p50 s (231)	p95 s (231)	hard max s	bundle	note
apus-decision-v1-4b	1.000	0.986	0.559	0.347	5.94	142.09	217.6	5.4G	three model servers on one GPU (30-item solo re-check: identical probabilities)
decider-0.8b	1.000	0.833	0.414	0.320	1.69	29.10	90.6	1.2G	server alone
laya-multilingual	0.896	0.403	0.342	0.273	0.02	0.20	0.3	649M	encoder s256: hard 95/111 rows truncated at 256 tokens
minicpm5-2b	0.979	0.708	0.459	0.241	0.14	1.19	1.8	2.5G	zero-shot chat, catalog temperature 2.93
openthai-systemone	1.000	0.819	0.324	0.449	0.62	25.93	39.0	1.0G	server alone (three short model loads overlapped; 5-item solo re-check identical)
qwen3.5-2b	1.000	0.750	0.441	0.245	2.74	50.22	62.3	2.8G	zero-shot chat, no calibration (temperature 1), server alone (kit gate runs overlapped 08:29–09:02)
qwen3.5-2b-decision	1.000	0.778	0.405	0.235	3.24	92.89	138.5	2.8G	three model servers on one GPU (30-item solo re-check: identical probabilities)
system-one-scorer-4b	1.000	0.861	0.505	0.164	8.96	61.61	72.0	4.7G	author max_len 384/row: hard 70/111 rows state cut; three model servers on one GPU (30-item solo re-check: identical probabilities)

published system (same 231 public items, their own serving stack)	easy	standard	hard	p50 s easy/std/hard
Jev 1.13.0 (TypeSafe AI)	1.000	0.986	0.730	0.665/0.653/0.672
Winnow-12B Q8	1.000	0.958	0.730	0.203/0.209/0.291
classifier.dev (fast tier)	1.000	0.986	0.703	0.378/0.387/0.374
decider-35b-a3b (Mapika)	1.000	0.972	0.667	0.283/0.283/0.312
SemIf, formerly OpenJev (Qwen3.5-4B, TheoLeeCJ)	1.000	0.986	0.613	0.188/0.188/0.223
reflex 4B (kshetrajna12)	1.000	0.944	0.604	0.33/1.128/1.879
open-alternative-jev (Qwen3.5-4B, IkerMoel)	1.000	0.833	0.568	0.191/0.192/0.247
decider-2b (Mapika)	1.000	0.847	0.495	0.26/0.259/0.268
system-one-open (Gemma 4 E2B LoRA on an L4)	1.000	0.931	0.486	0.516/0.649/0.682
kev 0.6B (research preview)	1.000	0.806	0.432	0.421/0.589/0.608
Laya (Convai Innovations, ModernBERT-large 421M)	0.958	0.694	0.351	0.368/0.382/1.922

Footnotes

laya-multilingual: encoder bundle wfp16-s256 → hard 95/111 rows truncated at 256 input tokens (states up to 3,677 tokens); easy/original untruncated (max 107).
system-one-scorer-4b: author's encode cuts the state from its end so each row (state + question + option) fits max_len 384 → hard 70/111 rows at the cap.
minicpm5-2b: kind chat, zero-shot under the kit's JSON decision prompt, catalog calibration temperature 2.93 (fit on SemIf perturbations108).
decider-0.8b / openthai-systemone / qwen3.5-2b-decision: Qwen3.5-family hybrids, S=1 prefill; a 3k-token hard state costs 40–220 s per question, re-prefilled for every question at this kit commit.
apus-decision-v1-4b hard p50: 44.0 s over the 55 rows answered while three serves ran, 8.4 s over the 56 rows answered alone.
Solo subset p50 (recheck30, easy / original / hard): qwen3.5-2b-decision 1.07 / 1.18 / 22.0 s; apus 2.60 / 2.62 / 41.9 s; scorer 2.51 / 2.73 / 20.1 s.
Published rows: recomputed from jevbench results/v1.2/jevbench-v1.2-per-task.json (per-item outcomes for the same 231 public ids); their runs used the entrants' own serving stacks (bf16 GPU etc.), so same items, not same conditions. JevBench's headline Intelligence/score is chance-corrected over 534 items incl. sealed ones and is not comparable to these accuracies.
qwen3.5-2b stands in for qwen3.5-4b, which is not in the kit catalog: kind chat, zero-shot under the kit's JSON decision prompt, no calibration entry (temperature 1).
Solo re-check (30 rows per model of the three-server run + 5 overlapped openthai rows): returned probabilities bit-identical to the concurrent runs (max |Δp| 0.0, argmax 30/30), latency ~2× lower alone.

In the qwen3.5-2b note, the kit gate runs were another session's decide-cli and swift build runs on the same Mac. They overlapped 73 and 42 of its rows; its last 67 hard rows ran with neither.

Method

Machine: Mac, Apple M4 Max, 128 GB, macOS 27.0 (26A428), Xcode 27.0 RC. Runs on 2026-09-24, JST (the manifests use UTC).
Server: decide-cli serve --model <catalog id> from coreai-kit adbc755 (Examples/Decide, release build) on 127.0.0.1. It answers the /v1/systemone requests that JevBench's typesafe adapter sends. Each model loads the bundle that the kit catalog pins at that commit (table below).
Client: JevBench 2fa63fa with the stock typesafe adapter and no API key. decider-0.8b ran through the JevBench CLI. The other seven ran through run_tier.py, which calls the same JevBench runner and summarizer with a 900 s HTTP timeout instead of the adapter's 120 s.
Per model: a 3-item contract probe, JevBench's conformance/check.py, then the easy, original (= standard) and hard tiers.
Scoring is JevBench's own: the predicted label is the argmax of the returned probabilities. All 8 × 231 requests returned HTTP 200, and every row was strict-valid.
verify.py recomputes accuracy, ECE (top label, 10 bins) and latency from the rows. Accuracy and the hard-tier ECE equal JevBench's summary-<tier>.json. table.py rebuilds both tables.
Latency is the client's wall clock per request. p50 and p95 are over all 231 requests of a model.
Published rows: published-public231.json, recomputed from JevBench's per-item outcomes (results/v1.2/jevbench-v1.2-per-task.json) for the same 231 item ids.
Solo re-check: recheck_select.py picked 30 items (recheck30.jsonl). The three models that had run as concurrent serves re-ran them alone, and so did the 5 openthai-systemone rows that overlapped other models' loads. recheck_compare.py wrote <model>/recheck/compare.json.
Cascade: cascade/ scores a device-to-hosted cascade from these rows and JevBench's published per-item outcomes. No new inference ran for it.

catalog id	Hugging Face repo	revision	bundle variant
apus-decision-v1-4b	mlboydaisuke/APUS-Decision-v1-4B-CoreAI	`3e946e1c4fafd9c909492bb81e8f400439560aee`	`gpu-pipelined-b2/apus_decision_v1_4b_decode_int8hu_block32_sym`
decider-0.8b	mlboydaisuke/decider-0.8b-CoreAI	`ff60ccf563556efbf83442e891e57d68a1840728`	`gpu-pipelined/decider_0_8b_decode_int8hu_block32_sym`
laya-multilingual	mlboydaisuke/Laya-Multilingual-CoreAI	`1175a4e6231fdfe8946e6566276f8d71eb8f02ef`	`macos/wfp16-s256`
minicpm5-2b	mlboydaisuke/MiniCPM5-2B-CoreAI	`f306590f991d66cad23af06bf72464519b65ba9f`	`int8`
openthai-systemone	mlboydaisuke/OpenThai-SystemOne-CoreAI	`e48583dab6c527fb87f47e074a5724a701c05600`	`gpu-pipelined/openthai_systemone_decode_int8lin`
qwen3.5-2b	mlboydaisuke/qwen3.5-2B-CoreAI	`3aa6c97d9545b9f7de6307b953a81db38cb660af`	`gpu-pipelined/qwen3_5_2b_decode_int8hu_block32_sym`
qwen3.5-2b-decision	mlboydaisuke/Qwen3.5-2B-Decision-CoreAI	`3b79912057db40be51697dfdaa433d6cb436b409`	`gpu-pipelined/qwen3_5_2b_decision_decode_int8hu_block32_sym`
system-one-scorer-4b	mlboydaisuke/system-one-qwen3.5-4b-scorer-CoreAI	`f27dcd3124cb3aa53e46b6707d3edb976f55c5f3`	`gpu-pipelined/system_one_qwen3_5_4b_scorer_decode_int8lin`

Caveats

Latency is contended and is reference only. The GPU was not reserved for these runs. qwen3.5-2b-decision, apus-decision-v1-4b and system-one-scorer-4b ran as three concurrent serves, and other jobs overlapped some rows of openthai-systemone and qwen3.5-2b (run.md lists them).
Concurrency changed latency, not answers. The solo re-check returned bit-identical probabilities: max |Δp| 0.0, argmax 30/30 for each of the three concurrent models and 5/5 for openthai-systemone. For the three concurrent models, the median solo/concurrent latency ratio was 0.43–0.52.
Truncation: the laya-multilingual bundle reads 256 input tokens, so 95 of 111 hard rows were cut. system-one-scorer-4b follows its author's 384-token rows, which cut the state on 70 of 111 hard rows.
qwen3.5-2b and minicpm5-2b are general chat models, asked zero-shot through the kit's JSON decision prompt.
Public items only. JevBench's official score combines four axes and includes sealed items, so it is not comparable with these accuracies. The published rows used each entrant's own serving stack: same items, different conditions.
The cascade chose and scored its thresholds on the same 231 items, so its accuracies are optimistic.
The numbers are for coreai-kit adbc755. The later 0.7.0 release was not re-run.

Layout

path	content
`RESULTS.md`	the two tables and the footnotes above
`published-public231.json`	per-tier accuracy and p50 of 52 published systems on the same 231 items
`recheck30.jsonl`	the 30 JevBench public items of the solo re-check, copied verbatim
`verify.py`, `table.py`	recompute one model's figures from its rows; rebuild the tables
`run_tier.py`	runs one tier through JevBench's runner and summarizer
`recheck_select.py`, `recheck_compare.py`	pick the re-check items; compare solo and concurrent rows
`cascade/`	`cascade.py`, `summary.json` and `README.md` of the cascade analysis
`LICENSE-jevbench`	JevBench's MIT licence, which covers the copied items
`<model>/results-{easy,original,hard}.jsonl`	one row per item, as JevBench's runner wrote it
`<model>/summary-{easy,original,hard}.json`	JevBench's summarize output per tier
`<model>/manifest-{easy,original,hard}.json`	adapter, endpoint, dataset hash, UTC start and end
`<model>/run.md`	versions, load, probe, conformance, timings and overlaps of the run
`<model>/serve.log`	one line per answered request: state tokens and milliseconds
`<model>/probe/results-probe.jsonl`	the 3-item contract probe (decider-0.8b: `probe/results.jsonl`)
`<model>/recheck/`	`results-recheck.jsonl` and `compare.json` of the solo re-check (4 models)

A result row holds the item id, family, the probability returned for each label, the predicted label, correctness, validity flags, HTTP status, latency and token usage. It holds no state or question text: grep -lE '"(state|question|instructions|criteria|expected|labels)":' */results-*.jsonl */probe/results*.jsonl */recheck/results-recheck.jsonl prints nothing. The item text lives in JevBench's datasets/public/; recheck30.jsonl is the only file here that copies it.

The scripts read a JevBench checkout at ~/code/jevbench. table.py also reads a coreai-kit checkout and the kit's local model store for its bundle column. Edit those paths to rerun them; verify.py needs only a model directory.

run.md also names local files that are not published: raw request and response bodies, other logs, process snapshots and ledgers. In this copy, the manifests' driver path is relative to the dataset root, and run.md omits local paths and the names of unrelated processes.

Attribution and license

The items, the harness and the published per-item outcomes come from JevBench by Florian Standhartinger and contributors, under the MIT licence (LICENSE-jevbench). published-public231.json is recomputed from JevBench's results/v1.2/jevbench-v1.2-per-task.json (file revision v1.3.0, at commit 2fa63fa). Each item copied into recheck30.jsonl carries the licence MIT in its own provenance field.

This dataset is MIT, like JevBench. It holds results and scripts, no model weights. Each model keeps the licence on its model card; system-one-scorer-4b, for example, is CC-BY-NC-4.0.