inclusionAI/AReno v0.0.3

v0.0.3

Repository: inclusionAI/AReno

Tag: v0.0.3

Published: 2026-06-29T12:53:00Z

Prerelease: no

Release notes:

Features

Agentic RL Examples

AReno v0.0.3 expands the agentic RL surface with runnable browser/game examples and training documentation. The release adds a DuelGrid browser demo with before/after training artifacts, a Tic-Tac-Toe agentic web UI, and a multi-turn coding agentic training example.

These examples are intended to make AReno's self-contained local agentic training flow easier to inspect: agent functions call AReno's local OpenAI-compatible proxy, return explicit trajectories, and train from the resulting messages, tool calls, rewards, logprobs, and loss masks.

Agentic Trajectory Controls

Agentic trajectories now trust tool calls parsed by the proxy instead of reparsing model text downstream. This keeps training data aligned with the actual serving/proxy behavior.

The agentic path also gains a max context length cap for training configuration, plus a DuelGrid reward fix that penalizes unspent energy so the demo better reflects task progress.

CLI and Chat Template Control

The train and serve CLIs now support disabling tokenizer chat-template thinking mode. This is useful for models and workflows where explicit thinking-mode control changes the rendered prompt or served chat behavior.

Fixes

• Fixed native attention prefill padding so V is padded to qk_head_dim.
• Fixed padded SFT loss masking so loss_mask is honored.
• Fixed SFT prompt and response budget handling so max_prompt_tokens and max_new_tokens are enforced separately.
• Fixed messages SFT fallback behavior so prior turns are not duplicated into the assistant training target.
• Fixed SFT loss normalization so gradients are weighted by global target tokens instead of microbatch count.
• Added a guardrail for SFT batches where all rows are filtered.
• Defined weight-sync vocabulary for future rollout/weight-sync design work.

Documentation and CI

• Added static docs publishing workflows for GitHub Pages and the inclusionAI asystem site.
• Added training-effect documentation and visual artifacts for the DuelGrid demo.

What's Changed

• feat(agentic): add tictactoe web ui by @xsuler in #105
• feat(agentic): add multi-turn coding example by @xsuler in #89
• ci(docs): publish static docs to asystem by @xsuler in #103
• ci(docs): publish static docs to gh-pages by @xsuler in #102
• fix(agentic): penalize unspent DuelGrid energy by @xsuler in #100
• feat(agentic): add max context length cap by @xsuler in ef04ec8
• feat(cli): add chat template thinking toggle by @xsuler in #98
• fix(sft): weight loss by global target tokens by @xsuler in #91
• fix(sft): enforce prompt and response budgets by @xsuler in #93
• fix(sft): honor padded loss masks by @xsuler in #92
• fix(sft): keep fallback chat deltas stable by @xsuler in #94
• fix(attn): pad native prefill values by @xsuler in #95
• fix(sft): fail when all rows are filtered by Thomas Weaver in c993605
• fix(agentic): trust proxy-parsed tool calls by @xsuler in #88
• feat(agentic): add DuelGrid browser demo and docs by @xsuler in #86

Full Changelog: https://github.com/inclusionAI/AReno/compare/v0.0.2...v0.0.3