NousResearch/atropos
1
0
Fork 0
mirror of https://github.com/NousResearch/atropos.git synced 2026-04-19 12:57:58 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2025-07-27 02:52:35 +00:00
parent a0979eb08e
commit 52b505296c
2 changed files with 243 additions and 121 deletions
Download patch file Download diff file Expand all files Collapse all files

4
environments/README.md
View file

@ -626,7 +626,7 @@ Please act as an impartial judge and evaluate the quality of the responses provi
**Evaluation Methodology:**
1. **Model Response Generation**: Generate response to Arena-Hard prompt using configured temperature/tokens
2. **Thinking Validation**: If thinking mode enabled, validate exactly one `<think></think>` pair and extract content after tags
3. **Dual-Round Judging**:
3. **Dual-Round Judging**:
- Round 1: Judge model response (A) vs GPT-4 baseline (B)
- Round 2: Judge GPT-4 baseline (A) vs model response (B)
4. **Score Combination**: Average the two judgment scores using Arena-Hard logic
@ -635,7 +635,7 @@ Please act as an impartial judge and evaluate the quality of the responses provi
**Reward Function:**
- **Training**: Scores range from -1.0 to 1.0 based on combined judgment results
- 1.0: Model response clearly better than baseline
- 0.0: Tie between model and baseline
- 0.0: Tie between model and baseline
- -1.0: Baseline clearly better than model response
- **Invalid Thinking**: Automatic 0.0 score for malformed `<think></think>` tags
- **Evaluation**: Converted to Arena-Hard winrate format (0.0 to 1.0)