Bad Performance of llama-2-70b in Humaneval Benchmark #3427
Description
I use command below to test llama-2-70b-hf model with humaneval benchmark
export CUDA_VISIBLE_DEVICES=0,1,2,3
export HF_ALLOW_CODE_EVAL="1"
python -m lm_eval \
--model hf \
--model_args pretrained=/home/shared/models/meta-llama/Llama-2-70b-hf/,parallelize=True \
--tasks humaneval \
--confirm_run_unsafe_code \
--show_config \
--log_samples \
--batch_size 1 \
--output_path ./evalbut the result is very unusual with only 0.0183 pass@1, here are my result and configuration
INFO:lm_eval.loggers.evaluation_tracker:Saving results aggregated
INFO:lm_eval.loggers.evaluation_tracker:Saving per-sample results for: humaneval
hf (pretrained=/home/shared/models/meta-llama/Llama-2-70b-hf/,parallelize=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1
| Tasks |Version| Filter |n-shot|Metric| |Value | |Stderr|
|---------|------:|-----------|-----:|------|---|-----:|---|-----:|
|humaneval| 1|create_test| 0|pass@1| |0.0183|± |0.0105|{
"results": {
"humaneval": {
"alias": "humaneval",
"pass@1,create_test": 0.018292682926829267,
"pass@1_stderr,create_test": 0.010496292269168305
}
},
"group_subtasks": {
"humaneval": []
},
"configs": {
"humaneval": {
"task": "humaneval",
"dataset_path": "openai/openai_humaneval",
"test_split": "test",
"doc_to_text": "{{prompt}}",
"doc_to_target": "{{test}}\ncheck({{entry_point}})",
"unsafe_code": true,
"description": "",
"target_delimiter": " ",
"fewshot_delimiter": "\n\n",
"num_fewshot": 0,
"metric_list": [
{
"metric": "def pass_at_k(references: list[str], predictions: list[list[str]], k: list[int] = None):\n global compute_\n assert k is not None\n if isinstance(k, int):\n k = [k]\n res = compute_.compute(\n references=references,\n predictions=predictions,\n k=k,\n )\n return res[0]\n",
"aggregation": "mean",
"higher_is_better": true,
"k": [
1
]
}
],
"output_type": "generate_until",
"generation_kwargs": {
"until": [
"\nclass",
"\ndef",
"\n#",
"\nif",
"\nprint"
],
"max_gen_toks": 1024,
"do_sample": false
},
"repeats": 1,
"filter_list": [
{
"name": "create_test",
"filter": [
{
"function": "custom",
"filter_fn": "<function build_predictions at 0x7f81e4c91120>"
}
]
}
],
"should_decontaminate": false,
"metadata": {
"version": 1.0,
"pretrained": "/home/shared/models/meta-llama/Llama-2-70b-hf/",
"parallelize": true
}
}
},
"versions": {
"humaneval": 1.0
},
"n-shot": {
"humaneval": 0
},
"higher_is_better": {
"humaneval": {
"pass_at_k": true
}
},
"n-samples": {
"humaneval": {
"original": 164,
"effective": 164
}
},
"config": {
"model": "hf",
"model_args": "pretrained=/home/shared/models/meta-llama/Llama-2-70b-hf/,parallelize=True",
"model_num_parameters": 68976648192,
"model_dtype": "torch.float16",
"model_revision": "main",
"model_sha": "",
"batch_size": "1",
"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": "7ddb2b1",
"date": 1764037905.2037978,
"pretty_env_info": "PyTorch version: 2.8.0+cu128\nIs debug build: False\nCUDA used to build PyTorch: 12.8\nROCM used to build PyTorch: N/A\n\nOS: Ubuntu 24.04.2 LTS (x86_64)\nGCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0\nClang version: Could not collect\nCMake version: version 4.1.0\nLibc version: glibc-2.39\n\nPython version: 3.12.3 (main, Feb 4 2025, 14:48:35) [GCC 13.3.0] (64-bit runtime)\nPython platform: Linux-5.15.0-91-generic-x86_64-with-glibc2.39\nIs CUDA available: True\nCUDA runtime version: Could not collect\nCUDA_MODULE_LOADING set to: LAZY\nGPU models and configuration: \nGPU 0: NVIDIA H20-3e\nGPU 1: NVIDIA H20-3e\nGPU 2: NVIDIA H20-3e\nGPU 3: NVIDIA H20-3e\nGPU 4: NVIDIA H20-3e\nGPU 5: NVIDIA H20-3e\nGPU 6: NVIDIA H20-3e\nGPU 7: NVIDIA H20-3e\n\nNvidia driver version: 550.144.03\ncuDNN version: Probably one of the following:\n/usr/lib/x86_64-linux-gnu/libcudnn.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.11.0\n/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.11.0\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: 52 bits physical, 57 bits virtual\nByte Order: Little Endian\nCPU(s): 240\nOn-line CPU(s) list: 0-239\nVendor ID: GenuineIntel\nModel name: Intel(R) Xeon(R) 6759P-C\nCPU family: 6\nModel: 173\nThread(s) per core: 2\nCore(s) per socket: 60\nSocket(s): 2\nStepping: 1\nBogoMIPS: 5200.00\nFlags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk amx_bf16 avx512_fp16 amx_tile amx_int8 arch_capabilities\nHypervisor vendor: KVM\nVirtualization type: full\nL1d cache: 5.6 MiB (120 instances)\nL1i cache: 7.5 MiB (120 instances)\nL2 cache: 240 MiB (120 instances)\nL3 cache: 640 MiB (2 instances)\nNUMA node(s): 2\nNUMA node0 CPU(s): 0-119\nNUMA node1 CPU(s): 120-239\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: Not affected\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; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS Not affected\nVulnerability Srbds: Not affected\nVulnerability Tsx async abort: Mitigation; TSX disabled\n\nVersions of relevant libraries:\n[pip3] Could not collect\n[conda] Could not collect",
"transformers_version": "4.57.1",
"lm_eval_version": "0.4.9.1",
"upper_git_hash": null,
"tokenizer_pad_token": [
"<unk>",
"0"
],
"tokenizer_eos_token": [
"</s>",
"2"
],
"tokenizer_bos_token": [
"<s>",
"1"
],
"eot_token_id": 2,
"max_length": 4096
}I want to know are there any additional configurations need to set before testing humaneval bench mark for llama2-70b-hf ? Please!