Skip to content

Releases

Ragent6 0.2.2

Choose a tag to compare

@weicj weicj released this 19 May 07:49

Ragent6 0.2.2 publishes the current Chinese adaptation score line: zh-CN 0.2.2.

Highlights:

  • Keeps the original public en-US 0.2.0 score line intact.
  • Publishes manifests/ragent6_0_2_2_zh_CN.json as the current Chinese prompt-set release.
  • Clarifies repository/package versioning so the latest public repository release is 0.2.2.
  • Includes native harness tolerance for reasoning-only responses and non-dict tool arguments seen in local OpenAI-compatible servers.

Validation run before release:

Ragent6 0.2.0 en-US: 60/60 (invalid=0)
Ragent6 0.2.2 zh-CN: 60/60 (invalid=0)
release_audit en-US 0.2.0: pass
release_audit zh-CN 0.2.2: pass
python3 -m compileall -q ragent6 scripts: pass

Score reports should state the exact score line and locale. Do not mix en-US 0.2.0, historical zh-CN 0.2.0/0.2.1, and current zh-CN 0.2.2 results in one leaderboard.

Ragent6 0.2.0

Choose a tag to compare

@weicj weicj released this 02 May 05:19

Ragent6 0.2.0 is the first public pre-1.0 score line for the deterministic local Agent benchmark.

Highlights:

  • 60 local execution cases across 6 dimensions.
  • Deterministic partial scoring, weighted to 100 points.
  • Strict pass count retained as auxiliary x/60 evidence.
  • No LLM judge and no online dependency in the benchmark core.
  • Dual locale score lines under the same methodology: en-US and zh-CN.
  • Locale results must not be mixed in one leaderboard.
  • README and README.zh-CN both include the zh-CN reference score panel.
  • The reference score panel now includes R1-R6 dimension breakdowns, each out of 10.

Validation:

  • make test passed.
  • en-US mock: 60/60, invalid=0.
  • zh-CN mock: 60/60, invalid=0.
  • release_audit.py passed for both locale manifests.
  • git diff --check passed.

Manifests:

  • en-US: manifests/ragent6_0_2_0_en_US.json
  • zh-CN: manifests/ragent6_0_2_0_zh_CN.json