Getting only 31% on the SWE-bench Verified instead of 40%. What am I doing wrong?
My setup:
- swe-agent github clone dated May 11 (https://github.com/SWE-agent/SWE-agent/commit/39e8ae3)
- config from https://github.com/SWE-bench/SWE-smith/blob/main/agent/swesmith_infer.yaml
- vllm for hosting the model:
python3 -m vllm.entrypoints.openai.api_server
--tensor_parallel_size $(echo $CUDA_VISIBLE_DEVICES | tr -cd ',' | wc -c | awk '{print $1+1}')
--enforce_eager
--gpu_memory_utilization 0.7
--enable-auto-tool-choice
--tool-call-parser hermes
--model $model_path
--tokenizer $model_path
--served-model-name $model_name
--rope-scaling '{"factor": 4.0, "original_max_position_embeddings": 32768, "rope_type": "yarn"}'
--enable_prefix_caching
--seed 41
--port $port > $vllm_log_file 2>&1 &
- swe-agent inference:
sweagent run-batch
--config config/swesmith_infer.yaml
--agent.model.name hosted_vllm/$model_name
--num_workers 10
--agent.model.temperature 0.0
--agent.model.api_base http://localhost:$port/v1/
--agent.model.per_instance_call_limit 75
--instances.type swe_bench
--instances.subset verified
--instances.split test
--instances.shuffle False >> $run_log_file 2>&1
I will be glad for any comments