{ "results": { "math_instruct": { "exact_match,none": 0.6327358367133324, "exact_match_stderr,none": 0.0052245703347459605, "alias": "math_instruct" }, "gsm8k_0shot_instruct": { "alias": " - gsm8k_0shot_instruct", "exact_match,none": 0.9052312357846853, "exact_match_stderr,none": 0.008067791560015407 }, "meta_math_0shot_instruct": { "alias": " - meta_math_0shot_instruct", "exact_match,none": 0.6276, "exact_match_stderr,none": 0.006837616441401548 }, "meta_math_hard_0shot_instruct": { "alias": " - meta_math_hard_0shot_instruct", "exact_match,none": 0.3806646525679758, "exact_match_stderr,none": 0.013349170720370741 } }, "groups": { "math_instruct": { "exact_match,none": 0.6327358367133324, "exact_match_stderr,none": 0.0052245703347459605, "alias": "math_instruct" } }, "group_subtasks": { "math_instruct": [ "meta_math_hard_0shot_instruct", "meta_math_0shot_instruct", "gsm8k_0shot_instruct" ] }, "configs": { "gsm8k_0shot_instruct": { "task": "gsm8k_0shot_instruct", "dataset_path": "json", "dataset_kwargs": { "data_files": "/home/jobuser/controlllm/inference/llm_eval_harness/additional_tasks/gsm8k/test.jsonl" }, "test_split": "train", "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc: dict) -> dict:\n # note that the orginal gsm8k data does not follow boxed format in the solution, so use the extracted answer in input_correct_responses directly as group truth here\n # compared to additional_tasks->math->utils.py, we used the boxed string lookup from solution as groundtruth\n out_doc = {\n \"problem\": doc[\"input_question\"],\n \"solution\": doc[\"solution\"] if \"solution\" in doc else doc[\"input_correct_responses\"][0],\n \"answer\": doc[\"input_correct_responses\"][0], # note that in meta eval dataset, input_correct_responses is: [ \"11\", \"11.\", \"11.0\", \"11.0.\", \"11.00\", \"11.00.\", \"11\", \"11.\", \"11.0\", \"11.0.\", \"11.00\", \"11.00.\" ], we take the first one for exact match. TODO: may need to normalize this to account for model's variations\n \"meta_target\": doc[\"input_correct_responses\"]\n }\n if getattr(doc, \"few_shot\", None) is not None:\n out_doc[\"few_shot\"] = True\n return out_doc\n processed_doc = dataset.map(_process_doc)\n # # take the first 10 to debug\n # processed_doc = processed_doc.select(range(10))\n return processed_doc\n", "doc_to_text": "def doc_to_text(doc: dict) -> str:\n input_final_prompts = doc[\"input_final_prompts\"][0]\n if llama_eval_instruction in input_final_prompts:\n input_final_prompts.replace(llama_eval_instruction, meta_math_instruction)\n return input_final_prompts\n", "doc_to_target": "answer", "process_results": "def process_results(doc: dict, results: List[str]) -> Dict[str, int]:\n candidates = results[0]\n\n # this is the implementation of \"is output correct\" from https://github.com/Kipok/NeMo-Skills\n # answer = extract_answer(candidates)\n\n # if is_output_correct(answer, doc[\"answer\"]):\n # retval = 1\n # else:\n # retval = 0\n\n # this is the implementation of \"is output correct\" from lm-evluation-harness(same from https://github.com/meta-llama/llama-recipes/blob/main/tools/benchmarks/llm_eval_harness/meta_eval/meta_template/math_hard/utils.py)\n last_boxed_string = last_boxed_only_string(candidates)\n if not last_boxed_string:\n # No boxed string found, so we can't evaluate\n return {\"exact_match\": 0}\n unnormalized_answer = remove_boxed(last_boxed_string)\n answer = normalize_final_answer(unnormalized_answer)\n\n if answer.strip() == (doc[\"answer\"].strip() if isinstance(doc[\"answer\"], str) else str(doc[\"answer\"])) or is_equiv(answer, str(doc[\"answer\"])):\n retval = 1\n else:\n retval = 0\n\n results = {\n \"exact_match\": retval,\n }\n return results\n", "description": "", "target_delimiter": " ", "fewshot_delimiter": "\n\n", "num_fewshot": 0, "metric_list": [ { "metric": "exact_match", "aggregation": "mean", "higher_is_better": true } ], "output_type": "generate_until", "generation_kwargs": { "until": [], "do_sample": false, "temperature": 0.0, "max_gen_toks": 5120 }, "repeats": 1, "should_decontaminate": false, "metadata": { "version": 1.0 } }, "meta_math_0shot_instruct": { "task": "meta_math_0shot_instruct", "dataset_path": "parquet", "dataset_kwargs": { "data_files": "/home/jobuser/controlllm/inference/llm_eval_harness/additional_tasks/math/joined_math.parquet" }, "test_split": "train", "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc: dict) -> dict:\n out_doc = {\n \"problem\": doc[\"input_question\"],\n \"solution\": doc[\"solution\"],\n \"answer\": normalize_final_answer(\n remove_boxed(last_boxed_only_string(doc[\"solution\"]))\n ),\n # \"answer\": doc[\"input_correct_responses\"][0] if isinstance(doc[\"input_correct_responses\"], list) else doc[\"input_correct_responses\"], # note that in meta eval dataset, input_correct_responses is: [ \"11\", \"11.\", \"11.0\", \"11.0.\", \"11.00\", \"11.00.\", \"11\", \"11.\", \"11.0\", \"11.0.\", \"11.00\", \"11.00.\" ], we take the first one for exact match. TODO: may need to normalize this to account for model's variations\n \"meta_target\": doc[\"input_correct_responses\"]\n }\n if getattr(doc, \"few_shot\", None) is not None:\n out_doc[\"few_shot\"] = True\n return out_doc\n processed_doc = dataset.map(_process_doc)\n # # take the first 10 to debug\n # processed_doc = processed_doc.select(range(10))\n return processed_doc\n", "doc_to_text": "def doc_to_text(doc: dict) -> str:\n return doc[\"input_final_prompts\"][0]\n", "doc_to_target": "answer", "process_results": "def process_results(doc: dict, results: List[str]) -> Dict[str, int]:\n candidates = results[0]\n\n # this is the implementation of \"is output correct\" from https://github.com/Kipok/NeMo-Skills\n # answer = extract_answer(candidates)\n\n # if is_output_correct(answer, doc[\"answer\"]):\n # retval = 1\n # else:\n # retval = 0\n\n # this is the implementation of \"is output correct\" from lm-evluation-harness(same from https://github.com/meta-llama/llama-recipes/blob/main/tools/benchmarks/llm_eval_harness/meta_eval/meta_template/math_hard/utils.py)\n last_boxed_string = last_boxed_only_string(candidates)\n if not last_boxed_string:\n # No boxed string found, so we can't evaluate\n return {\"exact_match\": 0}\n\n try:\n unnormalized_answer = remove_boxed(last_boxed_string)\n except AssertionError:\n # If remove_boxed fails due to an assertion error, return exact_match 0\n return {\"exact_match\": 0}\n\n answer = normalize_final_answer(unnormalized_answer)\n\n if answer.strip() == (doc[\"answer\"].strip() if isinstance(doc[\"answer\"], str) else str(doc[\"answer\"])) or is_equiv(answer, str(doc[\"answer\"])):\n retval = 1\n else:\n retval = 0\n\n results = {\n \"exact_match\": retval,\n }\n return results\n", "description": "", "target_delimiter": " ", "fewshot_delimiter": "\n\n", "num_fewshot": 0, "metric_list": [ { "metric": "exact_match", "aggregation": "mean", "higher_is_better": true } ], "output_type": "generate_until", "generation_kwargs": { "until": [], "do_sample": false, "temperature": 0.0, "max_gen_toks": 5120 }, "repeats": 1, "should_decontaminate": false, "metadata": { "version": 1.0 } }, "meta_math_hard_0shot_instruct": { "task": "meta_math_hard_0shot_instruct", "dataset_path": "parquet", "dataset_kwargs": { "data_files": "/home/jobuser/controlllm/inference/llm_eval_harness/additional_tasks/math_hard/joined_math_hard.parquet" }, "test_split": "train", "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc: dict) -> dict:\n out_doc = {\n \"problem\": doc[\"input_question\"],\n \"solution\": doc[\"solution\"],\n \"answer\": normalize_final_answer(\n remove_boxed(last_boxed_only_string(doc[\"solution\"]))\n ),\n # \"answer\": doc[\"input_correct_responses\"][0] if isinstance(doc[\"input_correct_responses\"], list) else doc[\"input_correct_responses\"], # note that in meta eval dataset, input_correct_responses is: [ \"11\", \"11.\", \"11.0\", \"11.0.\", \"11.00\", \"11.00.\", \"11\", \"11.\", \"11.0\", \"11.0.\", \"11.00\", \"11.00.\" ], we take the first one for exact match. TODO: may need to normalize this to account for model's variations\n \"meta_target\": doc[\"input_correct_responses\"]\n }\n if getattr(doc, \"few_shot\", None) is not None:\n out_doc[\"few_shot\"] = True\n return out_doc\n processed_doc = dataset.map(_process_doc)\n # # take the first 10 to debug\n # processed_doc = processed_doc.select(range(10))\n return processed_doc\n", "doc_to_text": "def doc_to_text(doc: dict) -> str:\n # remove the last \"<|start_header_id|>assistant<|end_header_id|>\\n\\n\"\n input_final_prompt = doc[\"input_final_prompts\"][0]\n # # Remove the last occurrence of the substring\n # generation_prompt = \"<|start_header_id|>assistant<|end_header_id|>\\n\\n\"\n # if input_final_prompt.endswith(generation_prompt):\n # input_final_prompt = input_final_prompt[:-len(generation_prompt)]\n\n return input_final_prompt\n", "doc_to_target": "answer", "process_results": "def process_results(doc: dict, results: List[str]) -> Dict[str, int]:\n candidates = results[0]\n\n # this is the implementation of \"is output correct\" from https://github.com/Kipok/NeMo-Skills\n # answer = extract_answer(candidates)\n\n # if is_output_correct(answer, doc[\"answer\"]):\n # retval = 1\n # else:\n # retval = 0\n\n # this is the implementation of \"is output correct\" from lm-evluation-harness(same from https://github.com/meta-llama/llama-recipes/blob/main/tools/benchmarks/llm_eval_harness/meta_eval/meta_template/math_hard/utils.py)\n last_boxed_string = last_boxed_only_string(candidates)\n if not last_boxed_string:\n # No boxed string found, so we can't evaluate\n return {\"exact_match\": 0}\n\n try:\n unnormalized_answer = remove_boxed(last_boxed_string)\n except AssertionError:\n # If remove_boxed fails due to an assertion error, return exact_match 0\n return {\"exact_match\": 0}\n\n answer = normalize_final_answer(unnormalized_answer)\n\n if answer.strip() == (doc[\"answer\"].strip() if isinstance(doc[\"answer\"], str) else str(doc[\"answer\"])) or is_equiv(answer, str(doc[\"answer\"])):\n retval = 1\n else:\n retval = 0\n\n results = {\n \"exact_match\": retval,\n }\n return results\n", "description": "", "target_delimiter": " ", "fewshot_delimiter": "\n\n", "num_fewshot": 0, "metric_list": [ { "metric": "exact_match", "aggregation": "mean", "higher_is_better": true } ], "output_type": "generate_until", "generation_kwargs": { "until": [], "do_sample": false, "temperature": 0.0, "max_gen_toks": 5120 }, "repeats": 1, "should_decontaminate": false, "metadata": { "version": 1.0 } } }, "versions": { "gsm8k_0shot_instruct": 1.0, "math_instruct": 1.0, "meta_math_0shot_instruct": 1.0, "meta_math_hard_0shot_instruct": 1.0 }, "n-shot": { "gsm8k_0shot_instruct": 0, "meta_math_0shot_instruct": 0, "meta_math_hard_0shot_instruct": 0 }, "higher_is_better": { "gsm8k_0shot_instruct": { "exact_match": true }, "math_instruct": { "exact_match": true }, "meta_math_0shot_instruct": { "exact_match": true }, "meta_math_hard_0shot_instruct": { "exact_match": true } }, "n-samples": { "meta_math_hard_0shot_instruct": { "original": 1324, "effective": 1324 }, "meta_math_0shot_instruct": { "original": 5000, "effective": 5000 }, "gsm8k_0shot_instruct": { "original": 1319, "effective": 1319 } }, "config": { "model": "ControlLLMWrapper", "model_args": { "pretrained": "/shared/models/llama3-10b-hf-checkpoint-sft-padding-openmath-train-concat-lerp-mse-loss-16-no-detach-v2/hf/checkpoint-250000", "dtype": "bfloat16", "tensor_parallel_size": 1, "gpu_memory_utilization": 0.6, "data_parallel_size": 8, "max_model_len": 8192, "cpu_offload_gb": 0, "enable_prefix_caching": false, "add_bos_token": true, "seed": 42, "register_model": "/shared/models/llama3-10b-hf-checkpoint-sft-padding-openmath-train-concat-lerp-mse-loss-16-no-detach-v2/hf/checkpoint-250000" }, "batch_size": 2, "batch_sizes": [], "device": null, "use_cache": null, "limit": null, "bootstrap_iters": 100000, "gen_kwargs": null, "random_seed": 0, "numpy_seed": 1234, "torch_seed": 1234, "fewshot_seed": 1234 }, "git_hash": null, "date": 1735431087.9255536, "pretty_env_info": "PyTorch version: 2.4.0+cu118\nIs debug build: False\nCUDA used to build PyTorch: 11.8\nROCM used to build PyTorch: N/A\n\nOS: CBL-Mariner/Linux (x86_64)\nGCC version: (GCC) 11.2.0\nClang version: Could not collect\nCMake version: version 3.21.4\nLibc version: glibc-2.35\n\nPython version: 3.10.14 (main, Jul 14 2024, 22:24:12) [GCC 11.2.0] (64-bit runtime)\nPython platform: Linux-5.15.148.2-2.cm2-x86_64-with-glibc2.35\nIs CUDA available: True\nCUDA runtime version: 11.8.89\nCUDA_MODULE_LOADING set to: LAZY\nGPU models and configuration: \nGPU 0: NVIDIA A100-SXM4-80GB\nGPU 1: NVIDIA A100-SXM4-80GB\nGPU 2: NVIDIA A100-SXM4-80GB\nGPU 3: NVIDIA A100-SXM4-80GB\nGPU 4: NVIDIA A100-SXM4-80GB\nGPU 5: NVIDIA A100-SXM4-80GB\nGPU 6: NVIDIA A100-SXM4-80GB\nGPU 7: NVIDIA A100-SXM4-80GB\n\nNvidia driver version: 535.129.03\ncuDNN version: Probably one of the following:\n/usr/lib/libcudnn.so.8.9.5\n/usr/lib/libcudnn_adv_infer.so.8.9.5\n/usr/lib/libcudnn_adv_train.so.8.9.5\n/usr/lib/libcudnn_cnn_infer.so.8.9.5\n/usr/lib/libcudnn_cnn_train.so.8.9.5\n/usr/lib/libcudnn_ops_infer.so.8.9.5\n/usr/lib/libcudnn_ops_train.so.8.9.5\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\n\nCPU:\nArchitecture: x86_64\nCPU op-mode(s): 32-bit, 64-bit\nAddress sizes: 48 bits physical, 48 bits virtual\nByte Order: Little Endian\nCPU(s): 256\nOn-line CPU(s) list: 0-255\nVendor ID: AuthenticAMD\nModel name: AMD EPYC 7763 64-Core Processor\nCPU family: 25\nModel: 1\nThread(s) per core: 2\nCore(s) per socket: 64\nSocket(s): 2\nStepping: 1\nFrequency boost: enabled\nCPU max MHz: 3529.0520\nCPU min MHz: 1500.0000\nBogoMIPS: 4900.41\nFlags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca\nVirtualization: AMD-V\nL1d cache: 4 MiB (128 instances)\nL1i cache: 4 MiB (128 instances)\nL2 cache: 64 MiB (128 instances)\nL3 cache: 512 MiB (16 instances)\nNUMA node(s): 8\nNUMA node0 CPU(s): 0-15,128-143\nNUMA node1 CPU(s): 16-31,144-159\nNUMA node2 CPU(s): 32-47,160-175\nNUMA node3 CPU(s): 48-63,176-191\nNUMA node4 CPU(s): 64-79,192-207\nNUMA node5 CPU(s): 80-95,208-223\nNUMA node6 CPU(s): 96-111,224-239\nNUMA node7 CPU(s): 112-127,240-255\nVulnerability Gather data sampling: Not affected\nVulnerability Itlb multihit: Not affected\nVulnerability L1tf: Not affected\nVulnerability Mds: Not affected\nVulnerability Meltdown: Not affected\nVulnerability Mmio stale data: Not affected\nVulnerability Retbleed: Not affected\nVulnerability Spec rstack overflow: Mitigation; safe RET, no microcode\nVulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp\nVulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected\nVulnerability Srbds: Not affected\nVulnerability Tsx async abort: Not affected\n\nVersions of relevant libraries:\n[pip3] flake8==7.1.1\n[pip3] flash-attn==2.6.3+cu118torch2.4cxx11abifalse\n[pip3] mypy-extensions==1.0.0\n[pip3] numpy==1.24.3\n[pip3] torch==2.4.0+cu118\n[pip3] torch-tb-profiler==0.4.1\n[pip3] torchsummary==1.5.1\n[pip3] torchvision==0.19.0+cu118\n[pip3] triton==3.0.0\n[conda] Could not collect", "transformers_version": "4.46.2", "upper_git_hash": null, "tokenizer_pad_token": [ "", "128256" ], "tokenizer_eos_token": [ "<|eot_id|>", "128009" ], "tokenizer_bos_token": [ "<|begin_of_text|>", "128000" ], "eot_token_id": 128009, "max_length": 8192 }