Published
Updated
Decision Model Factory prints DF-JA-Large at 67.63 on S1MB Japanese Task Avg against a public snapshot high of 60.05
Yuma Ichikawa posted Decision Model Factory on October 27, 2026, at 03:19 UTC. arXiv 2610.34956 prints DF-JA-Large at 67.63 on S1MB Japanese Task Avg against a fixed public snapshot high of 60.05. No public weights on the pages this desk opened. This desk did not run a build.
Yuma Ichikawa posted Decision Model Factory on October 27, 2026, at 03:19 UTC. The Japanese text says the company’s model reached the top score on a Japanese judgment bench, names a factory that can build Jev-like decision models for a site’s needs, and says a proprietary runner is fast on CPU. A follow-up at 03:24 UTC links arXiv 2610.34956. At 05:11 UTC he attached a benchmark still. The share image on this page is that still.
The paper, submitted 26 October 2026 at 07:23:08 UTC, is by Yuma Ichikawa at Fujitsu and RIKEN AIP. The abstract says five factory-built models improve on their backbones across Japanese decision tasks, compact deployment, and long-context passage ranking, and that three beat the strongest results in a fixed public Japanese comparison. A dedicated runner more than doubles serving speed on matched multi-question workloads. The HTML manuscript, opened October 27, defines Noul, Choice, and Score through an option-logit readout with LoRA on Gemma 4 backbones.
Table 1 prints S1MB Japanese Task Avg before and after specialization. DF-JA-Large moves from 58.99 to 67.63 on Gemma 4 31B. DF-Large moves from 58.99 to 66.98 on the same backbone. DF-Long moves from 51.58 to 63.57 on Gemma 4 26B-A4B. DF-Small moves from 31.97 to 52.96 on E4B. DF-JA-Small moves from 31.97 to 52.49 on E4B. The text says DF-JA-Large, DF-Large, and DF-Long all surpass the highest score of 60.05 in a fixed public snapshot of 34 complete evaluations retrieved on 18 October 2026, and that the small models exceed Saihai 8B at 49.77.
Table 2 prints DF-Long at S1MB English 63.55, Japanese 63.57, JQaRA nDCG@10 90.1 against Ruri v3 at 86.9, MIRACL Japanese 93.7, MLDR Japanese 98.9, embedded-instruction ignore rate 899 of 900 across three languages, and development ECE 0.024 on 600 decisions. Table 3 prints eight questions on one shared state at 29.4 ms for the decision runner against 71.7 ms on the general path, 2.44 times, on one NVIDIA B200 in BF16.
A Hugging Face search for DF-JA and Decision Model Factory on October 27 returned no matching public checkpoint. This desk did not run a build.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @yuma_1_or, October 27, 2026, 03:19 UTC, status 2114919838405775625. The Japanese text says the company's model reached the top score on a Japanese judgment bench, names Decision Model Factory as the base that can build Jev-like decision models for a site's needs, and says a proprietary runner is fast on CPU. Status 2114921248501412195, 03:24 UTC, links arXiv 2610.34956 and titles Decision Model Factory: Building Specialized Sovereign Decision Models. Status 2114948061231546454, 05:11 UTC, attaches a benchmark still. The share image on this page is that still.
arXiv 2610.34956, Yuma Ichikawa, Fujitsu and RIKEN AIP. Submitted 26 Oct 2026, 07:23:08 UTC. License arXiv.org perpetual non-exclusive. The abstract says five factory-built models improve on their backbones across Japanese decision tasks, compact deployment, and long-context passage ranking, and that three surpass the strongest results in a fixed public Japanese benchmark comparison. A dedicated runner more than doubles serving speed on matched multi-question workloads.
The HTML manuscript, opened October 27. Table 1 prints S1MB Japanese Task Avg before and after specialization: DF-JA-Large 58.99 to 67.63 on Gemma 4 31B, DF-Large 58.99 to 66.98 on the same backbone, DF-Long 51.58 to 63.57 on Gemma 4 26B-A4B, DF-Small 31.97 to 52.96 on E4B, and DF-JA-Small 31.97 to 52.49 on E4B. The text says DF-JA-Large, DF-Large, and DF-Long all surpass the highest score of 60.05 in a fixed public snapshot of 34 complete evaluations retrieved on 18 October 2026, and that the small models exceed Saihai 8B at 49.77. Table 2 prints DF-Long S1MB English 63.55, Japanese 63.57, JQaRA nDCG@10 90.1, MIRACL Japanese 93.7, MLDR Japanese 98.9, embedded-instruction ignore rate 899 of 900, and development ECE 0.024. Table 3 prints eight questions on one shared state at 29.4 ms for the decision runner against 71.7 ms on the general path, 2.44 times. Answer types in the paper are Noul, Choice, and Score. A Hugging Face search for DF-JA and Decision Model Factory on October 27 returned no matching public checkpoint. This desk did not run a build.
Compare
67.63 is S1MB Japanese Task Avg for DF-JA-Large after specialization. 60.05 is the fixed public snapshot high on that comparison. Do not merge 67.63 with Decision Index 0.4.1 Full scores, with d3-ultra's 70.18, or with Jeva's BANKING77 91.62.
52.96 is DF-Small after distillation. 49.77 is the Saihai 8B cell the paper prints. 29.4 ms is eight shared-state questions on the runner. 71.7 ms is the general path on that same table.
Terms
- Decision Model Factory
- Yuma Ichikawa's recipe pipeline for specialized decision models on Gemma 4 backbones. arXiv 2610.34956. Table 1 prints DF-JA-Large at 67.63 on S1MB Japanese Task Avg after specialization.
- S1MB Japanese Task Avg
- System One Mosaic Benchmark Japanese average in the paper: chance-corrected scores averaged within each answer type, then across Noul, Choice, and Score, times 100. DF-JA-Large prints 67.63 after specialization.
- DF-Long
- The factory's long-context specialist on Gemma 4 26B-A4B, about 4B active parameters per token. Table 2 prints S1MB Japanese 63.57 and JQaRA nDCG@10 90.1.