drop score very low

Hi there, thanks for the great work! I am trying to evaluate llama3.1-instruct on the tulu dev benchmarks with:

> olmes --model llama3.1-8b-instruct --task tulu_3_dev --output-dir llama_instruct_result --model-type vllm

But I only get a drop score of 0.14 (other benchmarks seems fine):

> {"task_name": "drop", "task_hash": "1f8759dfd5123aa42438dee687e84120", "model_hash": "77bf6fb4b5b1a783eedae6091247bb41", "model_config": {"model": "meta-llama/Meta-Llama-3.1-8B-Instruct", "revision": null, "trust_remote_code": null, "max_length": 8192, "model_path": null, "model_type": "vllm", "chat_model": true, "metadata": {"alias": "llama3.1-8b-instruct"}}, "task_config": {"task_name": "drop", "task_core": "drop", "limit": null, "split": "validation", "num_shots": 3, "fewshot_seed": 1234, "primary_metric": "f1", "random_subsample_seed": 1234, "context_kwargs": {"passage_prefix": "Passage: "}, "generation_kwargs": {"max_gen_toks": 50, "temperature": 0.0, "do_sample": false, "stop_sequences": ["Passage:", "Question:", "<|eot_id|>"]}, "metric_kwargs": null, "native_id_field": "query_id", "fewshot_source": null, "dataset_path": "EleutherAI/drop", "dataset_name": null, "use_chat_format": true, "version": 1, "revision": null, "metadata": {"regimes": ["Llama-3"], "alias": "drop::llama3"}}, "compute_config": {"batch_size": 1, "max_batch_size": 32, "output_dir": "llama_instruct_result", "num_recorded_inputs": 3, "save_raw_requests": true, "recompute_metrics": false}, "processing_time": 2372.8264021873474, "current_date": "2024-11-30 05:03:40 UTC", "num_instances": 9536, "beaker_info": {}, "metrics": {"f1": 0.1422934144295283, "exact_match": 0.03617869127516778, "primary_score": 0.1422934144295283}, "task_idx": 1}

Could you help me take a look? Thank you!