Decision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's "Decision Model"
Hacker News thread on Decision Index 0.2.1, which compared about 70 open reproductions of Jev's decision model.
This summary is based on how curated lists describe the project; its README was not available to review.
Overview
This Hacker News discussion is about Decision Index 0.2.1, an edition of the independent leaderboard for typed decision engines. The title describes it as covering 70 open reproductions of TypeSafe Jev's decision model.
The thread is useful for community reactions to how well open models replicate Jev and to the benchmark's methodology. The listing has no summary, and the thread content was not reviewed here.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Discussion of benchmark results for open decision models(inferred) | — |
How it works
- Read the thread alongside the Decision Index kit
- Compare open reproductions' scores on the leaderboard
- Reproduce results with the public Decision Index kit if needed
Strengths
- Community perspectives on open reproductions
- Context for how open reproductions compare
Caveats
- List-only; no summary in the record
- Older edition of the index
At a glance
- Type
- Discussion
- Status
- discussion
- Has code
- no
- Has tests
- no
Use cases and integrations
Frequently asked questions
What is discussed?
Decision Index 0.2.1 and its comparison of about 70 open Jev reproductions.
Where is the benchmark code?
In the apolinario/decision-index repository.
Is this the current edition?
No, the current board is 0.3.
Related projects
Bilingual research report on what Jev is, how its claims compare with independent evidence, and the open-source alternatives.
ProjectresearchAuto-generated write-up of a Reddit post claiming an open RLCD-trained Jev model; content is generic and largely inaccurate.
ProjectPythonDatawhale's Chinese Jev course: notebooks on the three primitives, translated recipes, evaluation and local Laya fine-tuning.
ArticleproductionPythonHead-to-head eval page where TrustedRouter's gev-1.0 outperforms Jev on the same JevBench questions and also accepts images.
ArticlediscussionAnthony Maio's Substack essay on what a model that cannot generate text is for, with a critical look at Jev's claims.
ArticlediscussionChinese X post rounding up five promising open Jev replicas: Laya 421M, Decider-2B, NanoJev 0.6B, Reflex and System-One 4B.
Discussiondiscussion