Added

• Added new models (OpenAI "0125" versions, Codellama, and more)
• Capability to evaluate code with AgentCodeFixer loop (set codefixing_num_rounds>0 )
• Automatically set a different seed for commercial API providers (MistralAI, OpenAI) to avoid their caching mechanism
• Re-scored all past submissions with the new methodology

Fixed

• Improved code loading and debugging via Julia's code loading mechanism (include_string), which allows to better locate the lines that caused the errors (run evaluate(....; verbose=true) to see which lines caused the errors or return_debug=true to return the debug information as a secondary output).
• Improved error capture and scoring (eg, imports of Base modules are now correctly recognized as "safe")
• Improved detection of parse errors (ie, reduces score of submissions that "executed" only because I didn't detect the parsing error earlier)
• Fixed mkdir bug in run_benchmark

Removed

• @timeout macro has been upstreamed to PromptingTools

Case Studies

• Quantization effects on Yi34b and Magicoder 7b
• Effect of English vs Chinese on performance with Yi34b

Added

• Documentation with detailed methodology, test case definitions, and results across various data cuts.
• Added ~5 samples for each model/prompt/test case combination for more robust results.