Skip to content
JevDirectory.orgTraffic count unavailable. View site traffic
Evaluation & TrainingTeacher-student distillation

Distillation

Training a smaller model to reproduce a larger model's output distribution, including its probabilities rather than labels alone.

All 164 terms
Category
Evaluation & Training
Also known as
Teacher-student distillation
Related terms
4
Directory entries
5
Docs
—
Added
2026-10-10

Definition

AutoTrust's JEV students are distilled from Jev 1.13 on a 741,000-row corpus; JEV-27B reports mean KL of about 0.017 to the teacher's distributions and choice top-1 agreement of 90.5% on held-out rows. The students inherit the teacher's failure modes, counting, dates, and arithmetic included, because the target is the teacher and not the truth.

Soft targets carry more information than hard labels, which is why distillation corpora store full distribution arrays and why teacher self-agreement becomes the ceiling. On the typed-decisions benchmark that ceiling is about 0.735, so scores above it mean learning teacher quirks.

Tagstrainingevaluation

Related terms

Definitions that connect to this one.

Proper scoring ruleA reward or error measure that is optimized only by reporting true probabilities — the training signal behind RLCD.Evaluation & TrainingExpected calibration errorThe average gap between predicted confidence and observed accuracy, grouped into bins. Lower is better, and the number depends on the bins.Evaluation & TrainingOpen weightsModel checkpoints published for anyone to download, run, and fine-tune, as opposed to a hosted closed API like Jev.EcosystemBenchmark claimsTypeSafe's headline multipliers — 193.6x faster, 444.6x cheaper, a 238x lower input price — come from its own workflow comparisons.Evaluation & Training

From the Jev directory

Jev Distill Corpus v3
The 741,000-row distillation corpus behind AutoTrust's Jev students: choice, score, and noul rows with full target distributions across yuri and openjev source streams, Apache-2.0.
Jev Benchmarks & EvaluationsDocs#community#evaluation#distillation
Community
AutoTrust JEV-27B
An open distillation of Jev 1.13 onto Qwen3.8-27B: a 108.9M-parameter decision block answers typed questions in one prefill pass and adds a POST /v1/decide route beside a full System 2.
Jev Models & Reimplementations#community#open-models#local
Community
Laya on Hugging Face
The Hugging Face hub for Laya's open System 1 checkpoints: the 421M English model at the repo root, a 322M multilingual build in a subfolder, and the typed-decisions fine-tune, all Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust JEV-9B
The 9B member of AutoTrust's Jev students: a 40.2M-parameter decision block on Qwen3.5-9B with 0.019 mean KL to Jev 1.13 and 70.7% HumanEval left untouched, small enough for one modest GPU.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust GEV-26B-Decide
AutoTrust's Gemma-based decision model with adaptive thinking: System 1 answers below 0.8 confidence hand off to the full Gemma 4 26B and mix the two by half, behind POST /v1/decide.
Jev Models & Reimplementations#community#open-models#local
Community

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai