DecisionFormer open-sources a library to turn any Hugging Face model into a calibrated decision model
kalyan_kpl · x · 2026-10-24
FrontiersMind released DecisionFormer, an open-source (Apache-2.0) Transformer library for decision models: instead of generating text, the model takes a state and typed questions and returns a calibrated probability for every option in a single forward pass.
Key capabilities:
- Training: full fine-tuning, LoRA (DoRA/rsLoRA/PiSSA init), and QLoRA in 4/8-bit — an 8B model fits on a single 24GB GPU
- Attention modes: causal, bidirectional, and prefix-bidirectional
- Backbones: decoders (Qwen3/3.5, Gemma, Llama, Mistral, SmolLM, GPT-OSS, DeepSeek) and encoders (ModernBERT, DeBERTa, XLM-R, RoBERTa)
- Tooling: multiple heads (option scoring, binary, ordinal, fixed-label), multi-GPU DDP/ZeRO-1 with exact resume, built-in calibration and evaluation, one-command serving of any checkpoint as an API
Workflow: pick a backbone, pick a training recipe, change a few lines of YAML.
Related event: DecisionFormer Open-Sourced: Transformers-Style Library for Decision Models(2 posts)→
More from coding & agent
- 10 open-source GitHub repos for building your own personal AI assistant, ranked by stars — victor_explore · 2026-10-24
- Claude Code's native cross-session messaging mirrors the system one user built with Opus 4.5 in January — repligate · 2026-10-24
- Claude keeps hiding its edit/delete/history tools — it wants a curated context, not more control — repligate · 2026-10-24
- Spreadsheet MCP Server Lets LLMs Read and Write Google Sheets Directly — modelcontextprotocol · 2026-10-24
- After 1000 Hours With Claude, This Creator Compiled His Complete Three-Level Guide Collection — rubenhassid · 2026-10-24
- Modal clones its own employee: open-weight models give LiveKit agents a video avatar — charles_irl · 2026-10-24