Fix weighted_summed_score in GEval metrics.

[SighingSnow]

A simple fix on issue #853 .

Original behavior:

For example, ('8', '0.01') and ('8', '0.001') are both in the score_logprobs,
After traversing ('8', '0.01'), token_linear_probability['8'] = 0.01, and sum_linear_probability=0.01
When coming to ('8', '0.001'), token_linear_probability['8'] = 0.001, and sum_linear_probability=0.011.

This will leads to the normalization error in the No.378 line.

Current behavior:

After traversing ('8', '0.01'), token_linear_probability['8'] = 0.01, and sum_linear_probability=0.01
When coming to ('8', '0.001'), token_linear_probability['8'] = 0.011, and sum_linear_probability=0.011.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
evals-docs	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Jun 26, 2024 3:55am

[penguine-ip]

@SighingSnow Great catch, TYSM!