PostHog/jeeves
Jeeves – Reasoning improves Jev-like decision models - PostHog/jeeves
186 stars · 9 forks
README
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).
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 (1.76×)
block 4, eight questions batched about 960 in total
Block 4 is the default because it stays cheap when several questions are batched. Reproduce Data You can build the datasets locally using the prep scripts. This downloads the public datasets from Hugging Face at the revisions pinned in prep/public.py: python -m prep.prep Each public dataset stays under its own license. Training On 8 GPUs, with the data in data/, bash run.sh runs the whole pipeline: torchrun --nproc_per_node 8 train.py sft --run-dir runs/sft torchrun --nproc_per_node 8 train.py cispo --run-dir runs/cispo --init runs/sft/final torchrun --nproc_per_node 8 test.py runs/cispo/final torchrun --nproc_per_node 8 jevbench.py runs/cispo/final python export.py runs/cispo/final --out runs/fused torchrun --nproc_per_node 8 -m drafter.gen --model runs/fused torchrun --nproc_per_node 8 train.py drafter --model runs/fused --block 4 --run-dir runs/drafter_k4 Repository
path contents
model/ Qwen3.5 (Gated DeltaNet + gated attention), LoRA, pointer head
loader/ prompt format, tokenisation and batching
prep/ dataset construction (prep.py) and synthetic generators
trainer.py, train.py SFT, CISPO and drafter training
test.py, jevbench.py, calibrate.py evaluation, JevBench, temperature fitting
export.py fuses LoRA into a standalone model with the head and temperature
drafter/ drafter model, chain sampling, fused speculative decoder
inference/ FP8 kernel, batched speculative engine, Jev-compatible server and benchmark
sdk/ jeeves_sdk, a drop-in replacement for Jev's Python SDK with the reasoning options
Limitations
Knowledge questions trail Jev (MMLU 0.793 vs 0.900, MMLU-Pro 0.739 vs 0.840). Thinking is slow at the tail: 17 s at p90 with full chains. Use max_think and nothink_threshold when latency matters. The Kev and Jev comparisons outside JevBench use different items from the same sources. No language consistency reward was included so thinking chains are not well interpretable.
Quote this If you use Jeeves, its training recipe or its drafter, please cite: @software{waltz2026jeeves, author = {Waltz, Nicholas P.}, title = {Jeeves: Reasoning Improves Jev-like Decisions}, year = {2026}, url = {https://github.com/PostHog/jeeves}, note = {Qwen3.5-9B decision model trained with SFT and CISPO, with a block-4 diffusion drafter} } References
Jev's Architecture Unmasked, the Jev design that Kev and Jeeves follow. Qwen Team. Qwen3.5-9B, the base model. Yang, Kautz, Hatamizadeh. Gated Delta Networks: Improving Mamba2 with Delta Rule. ICLR 2025. Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022. MiniMax. MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention. 2025. Introduces CISPO. Guo, Pleiss, Sun, Weinberger. On Calibration of Modern Neural Networks. ICML 2017. Temperature scaling. Orthrus, arXiv 2605.12825. The diffusion drafter ours adapts to Gated DeltaNet. Leviathan, Kalman, Matias. Fast Inference from Transformers via Speculative Decoding. ICML 2023.