Jeff vs Jev on the Same Questions: Overall Scores and a 26-Option Limit (Fixed in v1.1)

Jeff vs Jev on the Same Questions: Overall Scores and a 26-Option Limit (Fixed in v1.1)

Jeff is a new pair of open decision models, fine-tuned from Qwen3.5-0.8B and Qwen3.5-2B, that accept the same request format as Jev. Its README puts Jeff-2B just ahead of Jev, 83.1 to 83.0 over five public benchmarks, and marks Jeff as the winner of that row. It is also careful to say that Jev's figure "was measured on a different sample of the same benchmarks".

I have Jev API access and the same harness I used for Kev vs Jev, so I rebuilt Jeff's benchmark with Jeff's own script and ran Jev, Jeff-0.8B, Jeff-2B and Kev-9B on the same 4,599 questions. Then I ran Jeff on the six datasets from my earlier posts.

The short answer

• On the same questions, Jev was ahead: 85.7 against Jeff-2B's 83.0 (p < 0.001, paired). Jeff-2B was clearly better on Financial PhraseBank and RAGTruth, and clearly worse on BBH, JudgeBench and WinoGrande.
• The overall score depends on how the five benchmarks are combined. Jeff's "Overall" weights them by question count, a common choice, and two of them make up 54% of the questions. A plain mean of the five gives 79.2 against 85.5 on my run.
• In my tests Jeff v1.0 never chose an option past the 26th. On intent tasks with 60 to 150 options it scored 9–36%, while Kev-9B and Jev scored 62–83%. Reversing the option list moved the failures with it. Jeff's author fixed this in v1.1; the remeasured numbers are in the update below.

Reproducing Jeff's benchmark

Jeff's repository builds its benchmark with a script, jeff-panel, that downloads pinned revisions of five public datasets and draws a fixed-seed sample: 750 BBH questions (15 tasks, 50 each), 999 Financial PhraseBank sentences, 350 JudgeBench pairs, 1,500 RAGTruth responses and 1,000 WinoGrande sentences. I ran it unchanged. Before comparing anything, I checked that Jeff scored on my rebuild what its README reports:

Every figure is within 0.6 points, which is consistent with the questions and the serving setup being the same as Jeff's.

Jev on the same 4,599 questions

On these questions Jev scored 85.7, not the 83.0 in Jeff's README. The biggest change was Financial PhraseBank: 85.3 here against the 77.0 Jeff quotes. Jeff's panel uses the "all annotators agree" subset of Financial PhraseBank, the sentences on which every annotator gave the same sentiment, which is the easiest version of that dataset. That is a fair choice, but it means the 77.0 and the 85.3 are not the same test.

On the same questions Jeff-2B beat Jev on Financial PhraseBank by 11 points and on RAGTruth by 7. Jev was ahead on the other three: by 11 points on WinoGrande, 15 on JudgeBench and 23 on BBH. On BBH the gap was 23 points, and it was widest on the multi-step tasks: on reading a table of penguins Jev got 49 of 50 and Jeff-2B 20; on tracking shuffled objects, 47 against 24.

Training data matters for reading these numbers. Jeff's data sources list the training splits of RAGTruth and WinoGrande and a financial-news sentiment dataset of tweets, so three of the five panel benchmarks are close to tasks Jeff trained on (the panel items themselves were filtered out of training). TypeSafe does not publish what Jev was trained on, so I can't say the same for Jev either way.

Kev-9B, trained on none of these five, came in below both Jeff models overall. That fits what the Kev posts found: it does well on tasks shaped like its training data, and these reasoning and grounding benchmarks are not.

How benchmark weighting affects the overall score

Jeff's "Overall" is the share of all 4,599 questions answered correctly, so each benchmark counts in proportion to its size. That is a standard way to aggregate. The alternative, a plain mean of the five benchmark scores, gives each benchmark equal weight. The two can disagree here because the benchmarks differ in size: RAGTruth has 1,500 questions and Financial PhraseBank 999, 54% of the total, and those are Jeff's two strongest benchmarks.

Jev's published per-benchmark figures, weighted the same way, give exactly 83.0; averaged plainly, they give 83.6. Jeff-2B's published figures give 83.1 weighted and 79.4 averaged. So on the published figures the order depends on the aggregation. On the same questions, Jev is ahead under both.

Calibration: good overall, uneven by benchmark

Jeff's model card reports a calibration error of 0.028 for the 2B and 0.049 for the 0.8B. Pooling all 4,599 questions I got 0.028 and 0.046, so the card's figures appear to be pooled ones. By benchmark, Jeff-2B's error ranged from 1.2 points on RAGTruth to 14.6 on JudgeBench, where its average top probability was 13.3 points above its accuracy. Jev's ranged from 2.5 to 4.2 points on every benchmark, and 1.5 pooled.

A pooled number can look good while one benchmark is badly off, because errors in opposite directions cancel: Jeff-2B was over-confident on JudgeBench (+13.3 points) and WinoGrande (+6.9) and under-confident on Financial PhraseBank (−5.1). Using the same rule as the earlier posts, a threshold set so that accepted answers stay 95% correct would have accepted 54.4% of all questions for Jeff-2B and 63.0% for Jev. On JudgeBench it would have accepted 0.6% for Jeff-2B and 29.1% for Jev.

Off the benchmark: options past the 26th

Then I ran both Jeff models on the eight conditions from Kev vs Jev, same rows, same label names:

CLINC150's 100 out-of-scope questions have no correct option, so every model gets them wrong and 75% is the ceiling there.

The drop was on the three tasks with more than 26 options, and the failures showed a strong option-position pattern. Jeff labels options with letters, A to Z, then AA, AB and so on, and reads its answer as the probability of each label. In all 1,458 answers on those three tasks, neither Jeff model picked an option labelled past Z. When the right option was among the first 26, Jeff-0.8B got 80% right, about the same as Kev-9B (79%). Jeff-2B got 69%; the smaller model was also ahead on BANKING77 (23.4% against 16.2%), and I did not find out why. When it was 27th or later, each Jeff model got 0 of 452.

To check that position, not the questions, was the cause, I reversed the order of BANKING77's 77 options and ran Jeff-2B again. The 102 messages it had got 0 of now scored 25; the ones whose answers had moved past Z now scored 0. The right option past Z got a median probability of 0.0004. Jeff's README says a choice can have up to 255 options and that "the options can be anything: support queues, user intents, moderation labels, voice commands, game moves." The server accepts 255, but the models did not use any beyond the 26th here. MASSIVE is in Jeff's training data, and it still scored 35–36% there, because 102 of the 175 right answers sat past Z.

With v1.0, keep each question to 26 options or fewer, for example by shortlisting candidates in code first. On AG News both Jeff models were within four articles of Jev (p ≥ 0.29). On the six-option TREC task they trailed Jev by 13 to 26 points, so option position does not explain every gap.

Update, 30 September 2026: I reported this upstream, and Jeff's author confirmed the cause: the largest question in the training data had 19 options, so the two-letter codes were never trained. They released v1.1, retrained with lists of 20 to 254 options, and I reran the same tests on it. Both sizes now picked the right city at every position from 1 to 100. On BANKING77, which is still outside Jeff's training data, accuracy rose from 23.4% to 66.2% (0.8B) and from 16.2% to 64.9% (2B), against Jev's 76.0% on the same messages (p = 0.014 and 0.0015). CLINC150 and MASSIVE are now in Jeff's training data, so their new scores (72% and 83–84%) are no longer zero-shot. On Jeff's own benchmark v1.1 scored 79.1 (0.8B) and 81.9 (2B), in line with its README's 79.1 and 82.0; the 2B's small drop came mostly from JudgeBench (64.3 to 60.0).

SOTAAZ course

This course alone $10 · 23 lectures · every lecture freeWatch free →

Which to use

• Financial sentiment or checking a response against its source (RAGTruth-style): Jeff-2B was better than Jev on the same questions and runs locally.
• Reasoning, judging between two answers, pronoun resolution: Jev was 11 to 23 points ahead.
• More than 26 options: Jeff v1.1, not v1.0. On BANKING77's 77 options v1.1 scored 65–66%, against Jev's 76.0% and Kev-9B's 82.5%.

What this does not show

Jev is one version (jev-1.13.0). Jeff's panel is Jeff's own choice of benchmarks and samples; I kept it as is so the numbers compare with its README, and I did not run JevBench's hard tier or Jeff's game tests. Jeff was trained on the training splits of RAGTruth and WinoGrande, Kev on none of the five, and TypeSafe does not publish what Jev was trained on. The 26-option finding is from three intent datasets and one reversal test on one of them. I did not test Jeff-Gemma4-E2B.

The same 4,599-question panel was later run on Ollama 0.35's decision models: Nimble 9B scored 76.0 and Tev1 4B 70.8, against Jev's 85.7 and Jeff-2B v1.1's 81.9. Details in Ollama's decision models on the same questions as Jev.

The per-sentence results from this post were later re-read together with other decision models' results to collect the sentences many of them get wrong. Where Decision Models Go Wrong

Setup: Jeff repository firelex/jeff at 985797f; weights mstrasser/Jeff-Qwen3.5-0.8B 9878256 and mstrasser/Jeff-Qwen3.5-2B f2d993c, served with the repository's jeff-serve (PyTorch, bf16, torch 2.14.0) on one A100 80GB PCIe, 127.0.0.1 only, model name jeff-latest. Panel: jeff-panel unchanged, seed 20260926, 4,599 rows, none dropped for length, SHA-256 46b6fb82…. Kev-9B as in the earlier posts, served with kev.serve. Jev: POST https://api.typesafe.ai/v1/systemone with model: "jev-1.13.0", four requests at a time, no errors or retries. Every model got the panel's own state, instructions and criteria; yes/no questions (RAGTruth) are scored as "yes" at a probability of 0.5 or more. Calibration error is ECE over 10 bins of the top-option probability (for yes/no, max(p, 1 − p)); McNemar tests are exact and two-sided. The eight-condition runs use the same files, label names and instructions as the Kev posts. v1.1 remeasurement (2026-09-30): weights mstrasser/Jeff-Qwen3.5-0.8B 0f212b3 and Jeff-Qwen3.5-2B 2b1055e, code f067882, same harness and data; the position test is jeff-positions (100 options, 10 target cities × 2 shuffles, every position). The reversal test used the same 154 BANKING77 messages with the option list reversed, and every response's model field was jeff-qwen3.5-2b. Speed: Jeff answered the panel in a median of 45 ms per question on the A100 over local HTTP, for both sizes (Jeff's README reports 22–24 ms on an RTX PRO 6000, a setup I did not reproduce). The 4,599 Jev requests used 3.39 million input tokens, $0.14 at list price. Measured 2026-09-29.

All the text-classification measurements on this blog, put in the order you would use them: Classifier, LLM or Decision Model? A Measured Guide to Text Classification.