transformers merges 67-commit PR adding JEV + decision core, logic cut to templates + head loading
art_zucker · x · 2026-11-09
Hugging Face's transformers library merged PR #50054 (67 commits, by ArthurZucker) adding JEV + decision core to main. Key points: much of the logic previously living in remote code now boils down to chat templates plus head loading; DecisionHead classes were renamed to ForDecision with clef-only target capture moved out of core; decision heads now read the model passed to them, with decisionhead/ loaded from the decision repo, dropping cross-repo loading; configs use iscausal=False instead of usebidirectionalattention for qwen3.5 backbones; and encodemarkedchat computes its own marker pattern. The author hopes this helps people train better models.
Related event: Hugging Face transformers Adds Full Support for JEV Models(2 posts)→
More from Models
- Qwen Is the Workhorse: Nearly Every Decision Model on HuggingFace Fine-tunes From It — ivan_bezdomny · 2026-11-10
- LMArena launches Agent Arena with a Pareto frontier for agents — arena · 2026-11-10
- HuggingFace decision models are nearly all Qwen-based — ivan_bezdomny · 2026-11-10
- Kaggle Veteran: AI Benchmarks Need Private-Leaderboard-Style Hidden Test Sets — MeganRisdal · 2026-11-10
- Six free open-source models, one per job: from everyday chat to company servers — alex_verem · 2026-11-10
- Grok API text outputs now watermarked by default across 7 models for EU AI Act — thione · 2026-11-10