Jev 101
Start Guides
Search
中文 EN
Platform heat 2 · Cool★ 0heat incomplete (no X)

jev-gibbs

Replication of Large Language Gibbs (arXiv 2606.19264) with Jev as the per-claim conditional on TruthfulQA, plus controls showing most gain comes from context, not label consistency.

Repo: eumemic/jev-gibbs ↗

What it is

A careful research replication with Jev in the loop.

What it does: Scores TruthfulQA claims with Jev under zero-shot, context and Gibbs schedules.

Why it belongs in the showcase: Raw logs, controls and cost per method make the claims checkable.

Caveats: Research code; single model version.

Evidence

Public MIT repo (0★ at intake) created 2026-10-16 using jev-1.13.0. README results over the paper's 5 partitions: Jev zero-shot accuracy 0.913 for $0.005 per seed; context without verdicts (m=16) 0.931; block Gibbs m=16 0.931 with 2,240 calls. It reports the paper's best TruthfulQA result as 0.835. Includes analysis, switch logging, knights-and-knaves, entity resolution and kinship scripts with raw results.

Caveats

  • Research replication; numbers are author-reported on one model version
  • Some experiments use a rented GPU pod
  • Not a product; read the findings section before reuse

Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.