Matching Features, Not Tokens
Energy-Based Fine-Tuning (EBFT) is a language-model training method that matches activation-space features of generated and ground-truth completions using a frozen pretrained feature network. The project provides research documentation and code for machine learning researchers and developers experimenting with language-model fine-tuning.
Overview
7 featuresMatching Features, Not Tokens sits in PulseGate's Fine-tuning & training category. It focuses on reducing the distributional mismatch between language-model training data and the model's own generated completions. It is built as an open-source project for machine learning researchers and language model developers. The project is open source (Apache-2.0). Matching Features, Not Tokens is available on the web and the command line, and it can be self-hosted.
It is developed by Samy Jelassi, Mujin Kwun, Rosie Zhao, Yuanzhi Li, Nicolo Fusi, Yilun Du, Sham M. Kakade, Carles Domingo-Enrich, and it first shipped in 2026. The project is developed in the open on GitHub with 24 stars. Key capabilities include feature matching, on-policy sampling, and activation embeddings.
Summary written by a language model from the project’s public pages.
- Feature matching
- On-policy sampling
- Activation embeddings
- Frozen feature network
- Language-model fine-tuning
- Ground-truth comparison
- Reward-free training
Built with & integrations
- meta_llama
- llama- in the HTML
Trust & compliance
Indexing history
1What PulseGate has recorded for this listing
- Indexed4 Sept · 02:08 UTCMatching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models verified via Hacker News firehose (Algolia)Source: Hacker News firehose (Algolia) · Open
Frequently asked questions about Matching Features, Not Tokens
- What does Matching Features, Not Tokens do?
- Matching Features, Not Tokens focuses on reducing the distributional mismatch between language-model training data and the model's own generated completions. It is catalogued under Fine-tuning & training on PulseGate.
- Who is Matching Features, Not Tokens for?
- Matching Features, Not Tokens is an open-source project built for machine learning researchers and language model developers.
- Does Matching Features, Not Tokens have a free plan?
- Yes — Matching Features, Not Tokens is open source under the Apache-2.0 license and free to use.
- What platforms does Matching Features, Not Tokens run on?
- Matching Features, Not Tokens runs on the web and the command line. It can also be self-hosted.
- Is Matching Features, Not Tokens still maintained?
- Unverified. Matching Features, Not Tokens has not been re-checked since it entered the index, so there is no finding either way — and only a positive finding would say otherwise.
- What are alternatives to Matching Features, Not Tokens?
- Similar projects tracked by PulseGate include how-agentic, plato-neural-kernel, and annihilate-llm.how-agenticplato-neural-kernelannihilate-llm
- Who develops Matching Features, Not Tokens?
- Matching Features, Not Tokens is developed by Samy Jelassi, Mujin Kwun, Rosie Zhao, Yuanzhi Li, Nicolo Fusi, Yilun Du, Sham M. Kakade, Carles Domingo-Enrich.
- How long has Matching Features, Not Tokens been around?
- Matching Features, Not Tokens first shipped in 2026.
Also in Fine-tuning & training
Same category — not a similarity match