OR-decision-4B prints JevBench 193 of 231, and Choice 79.7 against Jev 79.6 on alphaXiv's chart

alphaXiv posted OR-decision-4B on October 30, 2026, at 16:06 UTC. The Hugging Face card prints JevBench 193 of 231, 83.55%. The launch chart prints Choice 79.7 against Jev 79.6, with sarcasm at 53.6 against Jev 70.7. Apache-2.0 LoRA on Qwen3.5-4B. This desk did not run the weights.

alphaXiv posted on October 30, 2026, at 16:06 UTC, that the lab trained a 4B decision model on over 117k decision-making examples with OpenResearch. The text says the model performs practically identically to Jev except on sarcasm detection, and that it generalized during evals to unseen tasks such as hate-speech severity, emotion detection, and intent classification. The share image on this page is that post’s chart.

Same-minute follow-ups add a Mammoth-1 comparison chart, a sarcasm bar, and links to alphaXiv/OR-decision-4B plus OpenResearch. NeuralLoot replied at 16:16 UTC that 53.6% on 140 sarcasm examples is near a coin flip if that subset is yes/no, and asked how much sarcasm sat in the 117k.

What the charts print

Opened October 31, the first chart prints Choice / Score / Yes/No / Sarcasm: Mammoth-1 85.6 / 69.1 / 91.0 / 62.1, Jev 79.6 / 65.4 / 87.4 / 70.7, OR-decision-4B 79.7 / 67.0 / 88.3 / 53.6, Elephant-1 74.7 / 62.3 / 84.7 / 60.0, Joey-1 71.0 / 61.8 / 81.1 / 57.1. A held-out task chart for OR-decision-4B prints Hate speech 72.1, Emotion 68.4, Intent 81.2, Toxicity 76.5, Sentiment 84.3, Topic 79.8, Urgency 71.6, Priority 73.9, Risk 69.2, and Compliance 77.4. The sarcasm chart reprints Jev 70.7 and OR-decision-4B 53.6.

What the card prints

The Hugging Face card, created October 30, 09:00 UTC and opened October 31, is Apache-2.0. It is a rank-16 LoRA on Qwen/Qwen3.5-4B. The corpus is 117,186 training examples after a content-hash split of 123,339 source examples. The evaluation table prints fixed validation 754 of 1,024, 73.63%; JevBench on a primary 4,096-token window 193 of 231, 83.55%; a secondary 3,072-token window 194 of 231, 83.98%; and base Qwen3.5-4B on the primary window 176 of 231, 76.19%. The JevBench pin is fstandhartinger/jevbench at 1bcc55eb6c8cffde2306b3db03ede39b61c6152a. The card says the Transformers/PEFT runtime has not been separately rebenchmarked. No hosted POST on the pages this desk opened. This desk did not run the weights.

83.55% is the card’s JevBench primary window. 79.7 is Choice on the launch chart. 53.6 is sarcasm on that chart. Those three stay on their own rows. Mammoth-1, Elephant-1, and Joey-1 stay on /news/mammoth-1/.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

alphaXiv, @askalphaxiv, October 30, 2026, 16:06 UTC, status 2116199951835582507. The text says the lab trained a 4B decision model on over 117k decision-making examples with OpenResearch, that the model performs practically identically to Jev in all areas except sarcasm detection, and that it generalized during evals to unseen tasks such as hate-speech severity, emotion detection, and intent classification. The share image on this page is that post's chart.

The same-minute follow-ups print a second chart with Mammoth-1, Jev, OR-decision-4B, Elephant-1, and Joey-1, then a sarcasm bar chart, then links to huggingface.co/alphaXiv/OR-decision-4B and github.com/alphaXiv/OpenResearch. NeuralLoot, @neuralloot, 16:16 UTC, status 2116202623519035670, wrote that 53.6% on 140 sarcasm examples is near a coin flip if that subset is yes/no, and asked how much sarcasm was in the 117k.

The first chart, opened October 31, prints Choice / Score / Yes/No / Sarcasm: Mammoth-1 85.6 / 69.1 / 91.0 / 62.1, Jev 79.6 / 65.4 / 87.4 / 70.7, OR-decision-4B 79.7 / 67.0 / 88.3 / 53.6, Elephant-1 74.7 / 62.3 / 84.7 / 60.0, Joey-1 71.0 / 61.8 / 81.1 / 57.1. The held-out task chart for OR-decision-4B prints Hate speech 72.1, Emotion 68.4, Intent 81.2, Toxicity 76.5, Sentiment 84.3, Topic 79.8, Urgency 71.6, Priority 73.9, Risk 69.2, Compliance 77.4. The sarcasm chart reprints Jev 70.7 and OR-decision-4B 53.6.

The Hugging Face card alphaXiv/OR-decision-4B, created October 30, 09:00 UTC, opened October 31. Apache-2.0. Rank-16 LoRA on Qwen/Qwen3.5-4B. Corpus 117,186 training examples after a content-hash split of 123,339 source examples. Evaluation table: fixed validation 754 of 1,024, 73.63%; JevBench primary 4,096-token window 193 of 231, 83.55%; secondary 3,072-token window 194 of 231, 83.98%; base Qwen3.5-4B on the primary window 176 of 231, 76.19%. JevBench pin is fstandhartinger/jevbench at 1bcc55eb6c8cffde2306b3db03ede39b61c6152a. The card says the Transformers/PEFT runtime has not been separately rebenchmarked. No hosted POST on the pages this desk opened. This desk did not run the weights.

Compare

83.55% is the card's JevBench primary window, 193 of 231. 79.7 is Choice on alphaXiv's launch chart against Jev 79.6 on that same chart. Do not merge 83.55 with 79.7.

53.6 is sarcasm on that chart, against Jev 70.7. NeuralLoot's reply names 140 sarcasm examples. It is not a JevBench composite and not Decision Index Full.

Mammoth-1 Choice 85.6 on the chart is not Base Labs' Decision Index 0.4 card Full 71.24. Joey-1 and Elephant-1 stay Base Labs rows on /news/mammoth-1/. Jeva-4B's BANKING77 91.62 is a different suite.

73.63% is the fixed validation set used for checkpoint selection. The card says that score is not an independent test estimate.

Terms

OR-decision-4B: alphaXiv's October 30, 2026 System One LoRA on Qwen3.5-4B, trained with OpenResearch on 117,186 examples. Hugging Face card prints JevBench 193 of 231, 83.55%. The launch chart prints Choice 79.7 against Jev 79.6, and sarcasm 53.6 against Jev 70.7. Apache-2.0.
OpenResearch: alphaXiv's MIT toolkit for training decision models with coding agents. Linked from the October 30 OR-decision-4B thread at github.com/alphaXiv/OpenResearch.