Jeeves. Reasoning improves Jev-like decision models

來源原標題

Jeeves. Reasoning improves Jev-like decision models

作者

nicowaltz

原始標籤

hackernews

來源訊號

HN points 212　HN 留言 85

原始連結

https://github.com/PostHog/jeeves

摘要原文

Jeeves – Reasoning improves Jev-like decision models A reasoning Jev-style classifier with a diffusion drafter, trained with SFT and CISPO. Acknowledgements Inspired by Kev . Highlights A 9B Jev-like model (Qwen3.5-9B, LoRA, pointer head) that thinks before it decides, with a block-4 diffusion drafter and the full training code and train/dev/test data. Beats Kev-9B and Jev on test data it was never trained on (0.889 vs 0.822 and 0.857) and on JevBench's public tiers (0.935 vs 0.866 for Jev). Supports yes/no ( noul ), multiple-choice ( choice ), and rating ( score ) questions in the same request, through a Jev-compatible API. About 0.3 s per request without thinking and a 3.3 s median with it on one H100. Can be sped up by truncating chain length. Runs on CUDA (Hopper for the FP8 kernel). Problem Jev-like models give calibrated decision probabilities, but at low accuracy. A lot of pipelines therefore rely on a reasoning model as a fallback. Jeeves trains a Jev-like Qwen3.5-9B (LoRA and a pointer head) using CISPO to reason before it decides. This results in better performance on out of domain tasks, and outperforms Jev in JevBench hard (public). Results Accuracy with thinking, greedy, 2,560-token cap. The Kev-9B and Jev columns are the numbers Kev publishes. bench Kev-9B Jev Jeeves Test overall (out-of-domain and held-out, item-weighted) 0.822 0.857 0.889 Transfer overall (MMLU-Pro and buried state) 0.579 0.800 0.746 JevBench overall (231 public items) 0.715* 0.866 0.935 QNLI 0.925 0.925 0.913 SciQ 0.963 0.988 0.991 TweetEval offensive 0.775 0.813 0.813 PAWS 0.763 0.788 0.875 MMLU 0.738 0.900 0.793 Emotion 0.600 0.588 0.647 Held-out rule structures 0.896 0.885 1.000 Contrastive policies 0.900 0.963 1.000 MMLU-Pro (10-way) 0.515 0.840 0.739 Buried state 0.740 0.700 0.759 Unknowable answered at p ≥ 0.9 (lower is better) 0.000 0.090 0.055 JevBench hard (111 public items) 0.451* 0.730 0.865 JevBench ECE (public items) 0.049 0.037 * No Kev-9B JevBench result is published. These are Kev-8B (Qwen3). All JevBench numbers are on the public easy, standard and hard tiers (231 items). The sealed judge tier is not included, and the Jev and Kev numbers are restricted to the same public items. Without thinking the same checkpoint scores 0.804 on our test split (2,962 items), against 0.840 with it. Quickstart Requirements: Python 3.12 and a CUDA GPU. pip install -r requirements.txt Download the released weights and serve them: hf download PostHog/jeeves --local-dir jeeves-weights python -m inference.serve --model jeeves-weights --drafter jeeves-weights/drafter_k4.safetensors --port 8009 Or fuse your own trained checkpoint into a standalone model and serve it with a drafter: python export.py runs/cispo/final --out runs/fused python -m inference.serve --model runs/fused --drafter runs/drafter_k4/drafter.safetensors --port 8009 Then send a request in Jev's format: curl -s localhost:8009/v1/systemone -H ' content-type: application/json ' -d ' { "state": "Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card.", "questions": { "department": {"type": "choice", "instructions": "Which team should handle this?", "criteria": {"returns": "Exchanges, refunds, wrong or damaged items", "shipping": "Delivery status, delays, lost packages", "billing": "Charges, invoices, payment problems"}}, "escalate": {"type": "noul", "instructions": "Does this need urgent human attention?"}, "frustration": {"type": "score", "instructions": "How frustrated is the customer?", "criteria": ["Calm", "Frustrated", "Very angry"]} }, "options": {"max_think": 512}} ' Response on one H100 (FP8), with the three questions thinking in parallel: { "model" : " jeeves-latest " , "answers" : { "department" : { "type" : " choice " , "choice" : " billing " , "confidence" : 0.19 , "probabilities" : { "returns" : 0.4 , "shipping" : 0.14 , "billing" : 0.46 } }, "escalate" : { "type" : " noul " , "noul" : 0.72 }, "frustration" : { "type" : " score " , "score" : 1.5 , "legend" : { "0" : " Calm " , "1" : " Frustrated " , "2" : " Very angry " }, "probabilities" : { "0" : 0.04 , "1" : 0.43 , "2" : 0.54 }, "confidence" : 0.75 } }, "usage" : { "input_tokens" : 129 , "output_tokens" : 160 , "reasoning_tokens" : 1536 }, "latency_ms" : 8141.6 } Python sdk/ is a drop-in replacement for Jev's Python SDK ( typesafe-sdk ): pip install ./sdk from jeeves_sdk import Choice , Noul , Score , TypeSafeClient with TypeSafeClient () as client : result = client . system_one ( state = "I was charged twice. Please help." , questions = { "billing" : Noul ( instructions = "Is this about billing?" ), "tone" : Choice ( instructions = "What is the tone?" , criteria = { "calm" : None , "angry" : None }), "urgency" : Score ( instructions = "How urgent is this?" , criteria = [ "can wait" , "this week" , "today" ]), }, max_think = 768 , return_reasoning = True , ) print ( result . nouls [ "billing" ]. noul , result . choices [ "tone" ]. choice , result . scores [ "urgency" ]. score ) print ( result . reasoning [ "tone" ]. text ) The client connects to http://127.0.0.1:8009 by default (or JEEVES_BASE_URL ), needs no API key, and waits up to 120s. Options options is optional and ignored by Jev clients that don't send it. Server-wide defaults are set with the matching serve flags. option default effect think true false answers from the prompt alone (about 0.3 s) max_think 2560 truncates each reasoning chain at this many tokens, then answers nothink_threshold null answers without thinking when the no-think confidence is at least this value return_reasoning false adds each question's reasoning text to the response On 325 dev questions: setting accuracy mean reasoning tokens median / p90 latency full thinking 0.825 1,138 3.3 s / 17.1 s max_think 768, nothink_threshold 0.9 0.806 344 2.0 s / 5.6 s no thinking 0.775 0 about 0.3 s How it works Questions, states and answers are loaded into the Qwen chat template like <state> …state… <q> instructions <opt> option 1 </opt> <opt> option 2 </opt> … <think> The model then rolls out its reasoning chain, and after the </think> token we append </think> <q> instructions <opt> option 1 </opt> <opt> option 2 </opt> … <decide> A pointer head scores each option with a scaled dot product between a query projection of the hidden state at <decide> and a key projection of the hidden state at that option's </opt> , where <state>, <q>, <opt>, </opt>, <decide> = "<|fim_prefix|>", "<|fim_middle|>", "<|box_start|>", "<|box_end|>", "<|fim_suffix|>" These are rare, largely unused tokens in the Qwen tokenizer. Ablations found that using plain text like "State" in the prompt instead worsened performance. Likewise, not repeating the questions after the reasoning block also decreases performance. The final probabilities are a softmax over the option scores, divided by a temperature fitted on the dev set. Training SFT (2 epochs, 596 steps on 8 GPUs). LoRA r=16 on all projections of Qwen3.5-9B plus the pointer head, trained on 19,126 questions from 12 public datasets and synthetic policy data. Half the questions carry a reasoning chain sampled from the base model. CISPO (a 624-step schedule stopped at step 402). 9,992 RL questions, 8 rollouts each at temperature 1, capped at 2,560 thinking tokens. Calibration . A single temperature fitted on dev, stored with the checkpoint. Stopping at step 402 keeps the best calibration and dev score. Past it, the head over-sharpens on the saturated RL pool. Diffusion drafter A diffusion view of the frozen model ( drafter/ ), inspired by Orthrus . Unlike Orthrus, which supports attention-only models, it supports Qwen3.5's Gated DeltaNet layers by letting mask tokens cross-attend to those layers' post-convolution keys and values. chain tokens per second plain graphed greedy decoding, one question 109 block 4, one question 176 (1.6×) block 8, one question 193 (1. (212 points, 85 comments on Hacker News)