| { |
| "repo_id": "yzeng58/benchpress-score-matrix", |
| "generated_at_utc": "2026-05-07T04:54:26.048511+00:00", |
| "source_git_commit": "5be3b4eddf0188721ff25f00713b589b2cbed8e0", |
| "source_git_branch": "yz", |
| "source_data_file": "benchpress/data/llm_benchmark_data.json", |
| "source_data_dirty": false, |
| "n_models_all": 189, |
| "n_benchmarks_all": 316, |
| "n_scores_all": 4903, |
| "paper_matrix": { |
| "m_threshold": 15, |
| "b_threshold": 8, |
| "iterations": 5, |
| "n_models": 84, |
| "n_benchmarks": 133, |
| "n_observations": 2604, |
| "fill_rate": 0.23308270676691728, |
| "dropped_models": [ |
| "amazon-nova-premier", |
| "amazon-nova-pro", |
| "claude-3-haiku", |
| "claude-3-opus", |
| "claude-3-sonnet", |
| "claude-3.5-haiku", |
| "claude-mythos", |
| "claude-sonnet-3.7", |
| "codestral-22b", |
| "command-r7b", |
| "deepseek-3.1", |
| "deepseek-r1-distill-llama-8b", |
| "deepseek-r1-distill-qwen-1.5b", |
| "deepseek-r1-distill-qwen-14b", |
| "deepseek-r1-distill-qwen-32b", |
| "deepseek-r1-distill-qwen-7b", |
| "deepseek-v3-0324", |
| "devstral-2", |
| "exaone-4.0-32b", |
| "exaone-deep-32b", |
| "falcon3-10b", |
| "gemini-1.5-flash-8b", |
| "gemini-2.0-flash-lite", |
| "gemini-2.0-pro", |
| "gemini-diffusion", |
| "gemma-2-2b", |
| "gemma-3-12b", |
| "gemma-3-4b", |
| "gemma-3n-e2b-instructed", |
| "gemma-3n-e2b-instructed-litert-preview", |
| "gemma-3n-e4b-instructed", |
| "gemma-3n-e4b-instructed-litert-preview", |
| "gemma-4-26b-a4b", |
| "gemma-4-31b", |
| "gemma-4-e2b", |
| "gemma-4-e4b", |
| "gpt-4-turbo", |
| "gpt-5.2-codex", |
| "gpt-5.3-codex", |
| "gpt-oss-20b", |
| "granite-3.3-8b-instruct", |
| "granite-4.0-1b", |
| "granite-4.0-h-1b", |
| "grok-1.5", |
| "grok-2", |
| "grok-2-mini", |
| "grok-4.20", |
| "ibm-granite-4.0-tiny-preview", |
| "internlm3-8b", |
| "lfm2.5-1.2b-instruct", |
| "llama-3.1-70b", |
| "llama-4-behemoth", |
| "llama-4-scout", |
| "longcat-flash-chat", |
| "minicpm-sala", |
| "minimax-m2.5", |
| "minimax-m2.7", |
| "mistral-large-2", |
| "mistral-large-3", |
| "mistral-medium-3", |
| "mistral-small-24b", |
| "mistral-small-3-24b-instruct", |
| "mistral-small-3.1", |
| "nemotron-ultra-253b", |
| "nova-lite", |
| "nova-micro", |
| "o1-mini", |
| "olmo-1b", |
| "olmo-2-13b", |
| "olmo-2-1b", |
| "olmo-2-32b", |
| "olmo-2-7b", |
| "olmo-7b-0424", |
| "openthinker2-32b", |
| "phi-3-14b", |
| "phi-3.5-mini-instruct", |
| "phi-3.5-moe-instruct", |
| "phi-4", |
| "phi-4-mini", |
| "phi-4-reasoning", |
| "phi-4-reasoning-plus", |
| "pixtral-12b", |
| "qwen-2.5-1.5b", |
| "qwen2-72b-instruct", |
| "qwen2-7b-instruct", |
| "qwen2.5-coder-32b-instruct", |
| "qwen2.5-coder-7b-instruct", |
| "qwen2.5-omni-7b", |
| "qwen2.5-vl-32b-instruct", |
| "qwen3-0.6b", |
| "qwen3-1.7b", |
| "qwen3-14b", |
| "qwen3-4b", |
| "sarvam-30b", |
| "seed-thinking-v1.5", |
| "smollm2-1.7b", |
| "tulu-3-8b" |
| ], |
| "dropped_benchmarks": [ |
| "aethercode", |
| "agentcompany", |
| "ai2d", |
| "aider_polyglot", |
| "ainstein_bench", |
| "all_angles", |
| "apex_agents", |
| "arc_challenge", |
| "arcagi1_image", |
| "arcagi2_image", |
| "babe", |
| "bfcl_v3_multiturn", |
| "bfcl_v4", |
| "bigbench_extra_hard", |
| "biobench", |
| "bixbench", |
| "blink", |
| "browsecomp_cm", |
| "browsecomp_long_context_256k", |
| "cgbench", |
| "chartqa", |
| "chartqapro", |
| "cl_bench", |
| "claw_eval_pass3", |
| "cmmlu", |
| "codesimpleqa", |
| "contphy", |
| "corpusqa_1m", |
| "countbench", |
| "covost2", |
| "creative_writing_v3", |
| "crossvid", |
| "ctf_internal", |
| "cybench", |
| "cybersecurity_ctf", |
| "da_2k", |
| "deepconsult", |
| "deepresearchbench", |
| "deepsearchqa_f1", |
| "der2_bench", |
| "disco_x", |
| "docvqa", |
| "dude", |
| "dynamath", |
| "egoschema", |
| "egotempo", |
| "emma", |
| "encyclo_k", |
| "eq_bench3", |
| "expert_swe", |
| "facts_benchmark", |
| "finsearchcomp", |
| "finsearchcompt23", |
| "flenqa_3k", |
| "fleurs", |
| "frontier_science_research", |
| "frontiersci_olympiad", |
| "frontiersci_research", |
| "fsc_147", |
| "gdpval_diamond", |
| "gdpval_oai_woe", |
| "genebench", |
| "graphwalks_bfs_128k_plus", |
| "graphwalks_bfs_256k_1m", |
| "graphwalks_parents_128k_plus", |
| "graphwalks_parents_256k_1m", |
| "hallusionbench", |
| "healthbench_consensus", |
| "hellaswag", |
| "hiddenmath", |
| "hipho", |
| "hle_verified", |
| "hle_vl", |
| "humaneval_plus", |
| "ib_modeling", |
| "imo_2025", |
| "infovqa", |
| "inverse_ifeval", |
| "korbench", |
| "livesports_3k", |
| "loft_128k", |
| "logicvista", |
| "longdocurl", |
| "longform_writing", |
| "longvideobench", |
| "lpfqa", |
| "lvbench", |
| "mars_bench", |
| "mathcanvas", |
| "mathkangaroo", |
| "mathvista_mini", |
| "medxpertqa_mm", |
| "medxpertqa_text", |
| "minedojo_verified", |
| "minerva", |
| "mm_browsecomp", |
| "mme_cc", |
| "mmlongbench", |
| "mmlongbench_doc", |
| "mmsibench_circular", |
| "mmstar", |
| "mmvu", |
| "morse_500", |
| "motionbench", |
| "mrcr_v2_2needle_1m", |
| "mrcr_v2_2needle_256k", |
| "mt_bench_101", |
| "mtvqa", |
| "muirbench", |
| "multilingual_mmlu", |
| "nl2repo_pass1", |
| "ntrex", |
| "ocrbench", |
| "ocrbench_v2", |
| "odvbench", |
| "officeqa", |
| "officeqa_pro", |
| "omnimath", |
| "ovbench", |
| "ovobench", |
| "paperbench", |
| "phibench", |
| "phybench", |
| "phyx_openended", |
| "point_bench", |
| "procbench", |
| "realworldqa", |
| "refspatialbench", |
| "researchrubrics", |
| "ruler_128k", |
| "ruler_32k", |
| "seal_0", |
| "sfe", |
| "simplevqa", |
| "spreadsheetbench_verified", |
| "superchem", |
| "swe_bench_multimodal", |
| "swe_evo", |
| "tau_bench_telecom", |
| "tempcompass", |
| "tob_complex_workflows", |
| "tob_compositional_tasks", |
| "tob_information_extraction", |
| "tob_k12_education", |
| "tob_referenceqa", |
| "tob_text_classification", |
| "tomato", |
| "treebench", |
| "tvbench", |
| "usamo_2026", |
| "videoeval_pro", |
| "videoholmes", |
| "videoreasonbench", |
| "videosimpleqa", |
| "visfactor", |
| "vispeak", |
| "visulogic", |
| "vitabench", |
| "viverbench", |
| "vlms_are_biased", |
| "vlms_are_blind", |
| "vpct", |
| "wildbench", |
| "world_travel_text", |
| "world_travel_vlm", |
| "worldvqa", |
| "xlrs_macro", |
| "xpert_bench", |
| "zerobench_main", |
| "zerobench_sub", |
| "zerobench_tools" |
| ], |
| "excluded_models": [ |
| "deepseek-v3.2-speciale", |
| "gemma-3-1b", |
| "granite-3.3-8b-base", |
| "kimi-k2-thinking", |
| "lfm2.5-1.2b-thinking", |
| "llama-3.2-1b", |
| "qwen-2.5-14b" |
| ], |
| "excluded_benchmarks": [ |
| "apex_shortlist", |
| "codeforces_avg8", |
| "codeforces_pass8", |
| "global_mmlu_lite", |
| "healthbench_hard", |
| "livecodebench_pro", |
| "livecodebench_v5", |
| "livecodebench_v6", |
| "mmlu", |
| "mmlu_redux", |
| "tau1_bench_avg", |
| "tau2_bench_avg" |
| ], |
| "model_canonical_rules": [ |
| { |
| "drop_id": "deepseek-v3.2-speciale", |
| "keep_id": "deepseek-v3.2", |
| "family": "DeepSeek-V3.2", |
| "reason": "same base model; keep the default-effort row because it has broader coverage" |
| }, |
| { |
| "drop_id": "kimi-k2-thinking", |
| "keep_id": "kimi-k2", |
| "family": "Kimi K2", |
| "reason": "same base model with a different mode; keep the broader-coverage default row" |
| }, |
| { |
| "drop_id": "lfm2.5-1.2b-thinking", |
| "keep_id": "lfm2.5-1.2b-instruct", |
| "family": "LFM2.5-1.2B", |
| "reason": "same base model with a different mode; keep the instruct/non-thinking row" |
| }, |
| { |
| "drop_id": "gemma-3-1b", |
| "keep_id": "gemma-3-1b-it", |
| "family": "Gemma 3 1B", |
| "reason": "same base model; many sources do not disambiguate base vs instruct, so use the base-only cells to fill missing instruct cells (small differences within evaluation noise)" |
| }, |
| { |
| "drop_id": "granite-3.3-8b-base", |
| "keep_id": "granite-3.3-8b-instruct", |
| "family": "Granite 3.3 8B", |
| "reason": "same base model; many sources do not disambiguate base vs instruct, so use the base-only cells to fill missing instruct cells (small differences within evaluation noise)" |
| }, |
| { |
| "drop_id": "llama-3.2-1b", |
| "keep_id": "llama-3.2-1b-instruct", |
| "family": "Llama 3.2 1B", |
| "reason": "same base model; many sources do not disambiguate base vs instruct, so use the base-only cells to fill missing instruct cells (small differences within evaluation noise)" |
| }, |
| { |
| "drop_id": "qwen-2.5-14b", |
| "keep_id": "qwen-2.5-14b-instruct", |
| "family": "Qwen2.5 14B", |
| "reason": "same base model; many sources do not disambiguate base vs instruct, so use the base-only cells to fill missing instruct cells (small differences within evaluation noise)" |
| } |
| ], |
| "benchmark_canonical_rules": [ |
| { |
| "drop_id": "codeforces_avg8", |
| "keep_id": "codeforces_rating", |
| "family": "Codeforces", |
| "reason": "same task family; keep Codeforces Rating because it has broader coverage" |
| }, |
| { |
| "drop_id": "codeforces_pass8", |
| "keep_id": "codeforces_rating", |
| "family": "Codeforces", |
| "reason": "same task family; keep Codeforces Rating because it has broader coverage" |
| }, |
| { |
| "drop_id": "livecodebench_pro", |
| "keep_id": "livecodebench", |
| "family": "LiveCodeBench", |
| "reason": "same benchmark family; keep the standard LiveCodeBench because it has broader coverage" |
| }, |
| { |
| "drop_id": "tau1_bench_avg", |
| "keep_id": "tau_bench_airline/tau_bench_retail", |
| "family": "τ-bench", |
| "reason": "composite average of existing domain benchmarks; keep per-domain benchmarks" |
| }, |
| { |
| "drop_id": "tau2_bench_avg", |
| "keep_id": "tau2_bench_airline/tau2_bench_retail/tau2_bench_telecom", |
| "family": "τ²-bench", |
| "reason": "composite average of existing domain benchmarks; keep per-domain benchmarks" |
| }, |
| { |
| "drop_id": "apex_shortlist", |
| "keep_id": "matharena_apex_2025", |
| "family": "MathArena Apex", |
| "reason": "subset/shortlist variant; keep MathArena Apex 2025 because it has broader coverage" |
| }, |
| { |
| "drop_id": "healthbench_hard", |
| "keep_id": "healthbench", |
| "family": "HealthBench", |
| "reason": "hard-subset variant; keep HealthBench because it has broader coverage" |
| }, |
| { |
| "drop_id": "mmlu", |
| "keep_id": "mmlu_pro", |
| "family": "MMLU", |
| "reason": "same benchmark family; keep MMLU-Pro because it has broader coverage" |
| }, |
| { |
| "drop_id": "mmlu_redux", |
| "keep_id": "mmlu_pro", |
| "family": "MMLU", |
| "reason": "same benchmark family; keep MMLU-Pro because it has broader coverage" |
| }, |
| { |
| "drop_id": "global_mmlu_lite", |
| "keep_id": "mmlu_pro", |
| "family": "MMLU", |
| "reason": "same benchmark family; keep MMLU-Pro because it has broader coverage" |
| }, |
| { |
| "drop_id": "livecodebench_v5", |
| "keep_id": "livecodebench", |
| "family": "LiveCodeBench", |
| "reason": "same benchmark family; use v5 only to fill missing standard LiveCodeBench cells, marking them non-canonical because the version differs" |
| }, |
| { |
| "drop_id": "livecodebench_v6", |
| "keep_id": "livecodebench", |
| "family": "LiveCodeBench", |
| "reason": "same benchmark family; use v6 only to fill missing standard LiveCodeBench cells, marking them non-canonical because the version differs" |
| } |
| ] |
| } |
| } |
|
|
