renderers: Token-Level Templating for Agentic RL
Open-sourcing renderers, a standalone Python library for token-level chat templating that keeps multi-turn RL rollouts stable, parseable, and packable.
renderers: Token-Level Templating for Agentic RL Today we're open-sourcing renderers , a standalone Python library that gives developers full control over conversation formatting for RL and multi-turn inference. The renderer abstraction, introduced by OpenAI's Harmony template for gpt-oss and popularized by Thinking Machines' Tinker cookbook , turns model chat templates into programmable Python objects. Instead of treating a chat template as a black-box Jinja string that formats messages, renderers exposes the operations RL systems actually need: rendering messages to token ids, parsing comple
saved by
related reading
- General Agent: A Self-Evolving, Synthetic Agent Environmentprimeintellect.ai
- General Agent: A Self-Evolving, Synthetic Agent Environmentprimeintellect.ai
- Recursive Language Models | Alex L. Zhangalexzhang13.github.io
- LLM Powered Autonomous Agents | Lil'Loglilianweng.github.io
- Keep the Tokens Flowing: Lessons from 16 Open-Source RL Librarieshuggingface.co
- Patterns for Building LLM-based Systems & Productseugeneyan.com
- Prime Agent: A Self-Improving RLM Harnessarxiv.org
- Tinkerthinkingmachines.ai
- Prime Agent: A self-improving RLM agentprimeintellect.ai
- State of RL for reasoning LLMs | A. Weersaweers.de
- A Taxonomy of RL Environments for LLM Agentsleehanchung.github.io
- GitHub - brexhq/prompt-engineering: Tips and tricks for working with Large Language Models like OpenAI's GPT-4.github.com