feat(evaluation): Added INSTRUCTION_TEMPLATE_NAME to run_infer.py in swe_bench (#10270)

Co-authored-by: Xingyao Wang <xingyao@all-hands.dev>
Co-authored-by: mamoodi <mamoodiha@gmail.com>
This commit is contained in:
Kevin Musgrave
2025-08-18 07:18:08 -07:00
committed by GitHub
parent bef6b1afee
commit 74ba21bad0
3 changed files with 9 additions and 1 deletions

View File

@@ -93,6 +93,9 @@ export USE_HINT_TEXT=true # Ignore this if you are not sure.
# Specify a condenser configuration for memory management (default: NoOpCondenser)
export EVAL_CONDENSER=summarizer_for_eval # Name of the condenser config group in config.toml
# Specify the instruction prompt template file name
export INSTRUCTION_TEMPLATE_NAME=swe_custom.j2 # Name of the file in the swe_bench/prompts folder.
```
Let's say you'd like to run 10 instances using `llm.eval_gpt4_1106_preview` and CodeActAgent,

View File

@@ -108,7 +108,9 @@ def get_instruction(instance: pd.Series, metadata: EvalMetadata) -> MessageActio
llm_model = metadata.llm_config.model
# Determine the template file based on mode and LLM
if mode.startswith('swt'):
if metadata.instruction_template_name:
template_name = metadata.instruction_template_name
elif mode.startswith('swt'):
template_name = 'swt.j2'
elif mode == 'swe':
if 'gpt-4.1' in llm_model:
@@ -122,6 +124,7 @@ def get_instruction(instance: pd.Series, metadata: EvalMetadata) -> MessageActio
logger.error(f'Unexpected evaluation mode: {mode}. Falling back to default.')
template_name = 'swe_default.j2'
logger.debug(f'Using instruction template file: {template_name}')
# Set up Jinja2 environment
# Assuming templates are in 'evaluation/benchmarks/swe_bench/prompts' relative to this script
prompts_dir = os.path.join(os.path.dirname(__file__), 'prompts')