lm-eval v0.4.9 Release Notes

Key Improvements

• Enhanced Backend Support:

  • SGLang Generate API by @baberabb in #2997
  • vLLM enhancements: Added support for enable_thinking argument (#2947) and data parallel for V1 (#3011) by @anmarques and @baberabb
  • Chat template improvements: Extended vLLM chat template support (#2902) and fixed HF chat template resolution (#2992) by @anmarques and @fxmarty-amd

• Multimodal Capabilities:

  • Audio modality support for Qwen2 Audio models by @artemorloff in #2689
  • Image processing improvements: Added resize images support (#2958) and enabled multimodal API usage (#2981) by @artemorloff and @baberabb
  • ChartQA multimodal task implementation by @baberabb in #2544

• Performance & Reliability:

  • Quantization support added via quantization_config by @jerryzh168 in #2842
  • Memory optimization: Use yaml.CLoader for faster YAML loading by @giuliolovisotto in #2777
  • Bug fixes: Resolved MMLU generative metric aggregation (#2761) and context length handling issues (#2972)

New Benchmarks & Tasks

Code Evaluation

• HumanEval Instruct - Instruction-following code generation benchmark by @baberabb in #2650
• MBPP Instruct - Instruction-based Python programming evaluation by @baberabb in #2995

Language Modeling

• C4 Dataset Support - Added perplexity evaluation on C4 web crawl dataset by @Zephyr271828 in #2889

Long Context Benchmarks

• RULER and Longbench - Long-context evaluation suites added by @baberabb in #2629

Mathematical & Reasoning

• GSM8K Platinum - Enhanced mathematical reasoning benchmark by @Qubitium in #2771
• MastermindEval - Logic reasoning evaluation by @whoisjones in #2788
• JSONSchemaBench - Structured output evaluation by @Saibo-creator in #2865

Llama Reference Implementations

• Llama Reference Implementations - Added task variants for Multilingual MMLU, MMLU CoT, GSM8K, and ARC Challenge based on Llama evaluation standards by @anmarques in #2797, #2826, #2829

Multilingual Expansion

Asian Languages:

• Korean MMLU (KMMLU) multiple-choice task by @Aprilistic in #2849
• MMLU-ProX extended evaluation by @heli-qi in #2811
• KBL 2025 Dataset - Updated Korean benchmark evaluation by @abzb1 in #3000

European Languages:

• NorEval - Comprehensive Norwegian benchmark by @vmkhlv in #2919

African Languages:

• AfroBench - Multi-African language evaluation by @JessicaOjo in #2825
• Darija tasks - Moroccan dialect benchmarks (DarijaMMLU, DarijaHellaSwag, Darija_Bench) by @hadi-abdine in #2521

Arabic Languages:

• Arab Culture task for cultural understanding by @bodasadallah in #3006

Domain-Specific Benchmarks

• CareQA - Healthcare evaluation benchmark by @PabloAgustin in #2714
• ACPBench & ACPBench Hard - Automated code generation evaluation by @harshakokel in #2807, #2980
• INCLUDE tasks - Inclusivity evaluation suite by @agromanou in #2769
• Cocoteros VA dataset by @sgs97ua in #2787

Social & Bias Evaluation

• Various social bias tasks for fairness assessment by @oskarvanderwal in #1185

Technical Enhancements

• Fine-grained evaluation: Added --examples argument for efficient multi-prompt evaluation by @felipemaiapolo and @mirianfsilva in #2520
• Improved tokenization: Better handling of add_bos_token initialization by @baberabb in #2781
• Memory management: Enhanced softmax computations with softmax_dtype argument for HFLM by @Avelina9X in #2921

Critical Bug Fixes

• Collating Queries Fix - Resolved error with different continuation lengths that was causing evaluation failures by @ameyagodbole in #2987
• Mutual Information Metric - Fixed acc_mutual_info calculation bug that affected metric accuracy by @baberabb in #3035

Breaking Changes & Important Updates

• MMLU dataset migration: Switched to cais/mmlu dataset source by @baberabb in #2918
• Default parameter updates: Increased max_gen_toks to 2048 and max_length to 8192 for MMLU Pro tests by @dazipe in #2824
• Temperature defaults: Set default temperature to 0.0 for vLLM and SGLang backends by @baberabb in #2819

We extend our heartfelt thanks to all contributors who made this release possible, including 43 first-time contributors who brought fresh perspectives and valuable improvements to the evaluation harness.

What's Changed

• fix mmlu (generative) metric aggregation by @wangcho2k in #2761
• Bugfix by @baberabb in #2762
• fix verbosity typo by @baberabb in #2765
• docs: Fix typos in README.md by @ruivieira in #2778
• initialize tokenizer with add_bos_token by @baberabb in #2781
• improvement: Use yaml.CLoader to load yaml files when available. by @giuliolovisotto in #2777
• Consistency Fix: Filter new leaderboard_math_hard dataset to "Level 5" only by @perlitz in #2773
• Fix for mc2 calculation by @kdymkiewicz in #2768
• New healthcare benchmark: careqa by @PabloAgustin in #2714
• Capture gen_kwargs from CLI in squad_completion by @ksurya in #2727
• humaneval instruct by @baberabb in #2650
• Update evaluator.py by @zhuzeyuan in #2786
• change piqa dataset path (uses parquet rather than dataset script) by @baberabb in #2790
• use verify_certificate flag in batch requests by @daniel-salib in #2785
• add audio modality (qwen2 audio only) by @artemorloff in #2689
• Add various social bias tasks by @oskarvanderwal in #1185
• update pre-commit by @baberabb in #2799
• Update Legacy OpenLLM leaderboard to use "train" split for ARC fewshot by @Avelina9X in #2802
• Add INCLUDE tasks by @agromanou in #2769
• Add support for token-based auth for watsonx models by @kiersten-stokes in #2796
• add version by @baberabb in #2808
• Add cocoteros_va dataset by @sgs97ua in #2787
• Add MastermindEval by @whoisjones in #2788
• Add loncxt tasks by @baberabb in #2629
• [hf-multimodal] pass kwargs to self.processor by @baberabb in #2667
• [MM] Chartqa by @baberabb in #2544
• Allow writing config to wandb by @ksurya in #2736
• [change] group -> tag on afrimgsm, afrimmlu, afrixnli dataset by @jd730 in #2813
• Clean up README and pyproject.toml by @kiersten-stokes in #2814
• Llama3 mmlu correction by @anmarques in #2797
• Add Markdown linter by @kiersten-stokes in #2818
• Configure the pad tokens for Qwen when using vLLM by @zhangruoxu in #2810
• fix typo in humaneval by @baberabb in #2820
• default temp=0.0 for vllm and slang by @baberabb in #2819
• Fixes to mmlu_pro_llama by @anmarques in #2816
• Add MMLU-ProX task by @heli-qi in #2811
• Quick fix for mmlu_pro_llama by @anmarques in #2827
• Fix: tj-actions/changed-files is compromised by @Tautorn in #2828
• Multilingual MMLU for Llama instruct models by @anmarques in #2826
• bbh - changed dataset to parquet version by @baberabb in #2845
• Fix typo in longbench metrics by @djwackey in #2854
• Add kmmlu multiple-choice(accuracy) task #2848 by @Aprilistic in #2849
• Adding ACPBench task by @harshakokel in #2807
• add Darija (Moroccan dialects) tasks including darijammlu. darijahellaswag and darija_bench by @hadi-abdine in #2521
• Increase default max_gen_toks to 2048 and max_length to 8192 for MMLU Pro tests by @dazipe in #2824
• doc by @baberabb in #2857
• Fix: ACPBench Link by @harshakokel in #2860
• Adds MMLU CoT, gsm8k and arc_challenge for llama instruct by @anmarques in #2829
• [leaderboard] math - sync with repo by @baberabb in #2817
• Update supported models by @danielholanda in #2866
• Add JSONSchemaBench: A Benchmark for Evaluating Structured Output from LLMs by @Saibo-creator in #2865
• leaderboard - add subtask scores by @baberabb in #2867
• Fix the deps of longbench from jeiba to jieba by @houseroad in #2873
• Optimization for evalita-llm rouge computation by @m-resta in #2878
• Update authentications methods, add support for deployment_id for IBM watsonx_ai by @Medokins in #2877
• Add GSM8K Platinum by @Qubitium in #2771
• Add --examples Argument for Fine-Grained Task Evaluation in lm-evaluation-harness. This feature is the first step towards efficient multi-prompt evaluation with PromptEval [1,2] by @felipemaiapolo in #2520
• Extend support for chat template in vLLM by @anmarques in #2902
• tasks README: fix dead link by @dtrifiro in #2899
• Add support for quantization_config by @jerryzh168 in #2842
• Fix a typo in README for tasks by @eldarkurtic in #2910
• fix resolve_hf_chat_template version by @baberabb in #2917
• mmlu - switch dataset to cais/mmlu; fix tests by @baberabb in #2918
• init pixels before tokenizer creation by @artemorloff in #2911
• Longbench bugfix by @baberabb in #2895
• Added softmax_dtype argument to HFLM to coerce log_softmax computations by @Avelina9X in #2921
• [bbh] use np.nan for numpy > 2.0 by @baberabb in #2937
• Add support for enable_thinking argument in vllm model by @anmarques in #2947
• Added NorEval, a novel Norwegian benchmark by @vmkhlv in #2919
• Fix import error for eval_logger in score utils by @annafontanaa in #2940
• Include all test files in sdist by @booxter in #2634
• Change citation name by @StellaAthena in #2956
• [vllm] add warning on truncation by @baberabb in #2962
• fix: type error while checking context length by @llsj14 in #2972
• Fix import error for deepcopy by @kiersten-stokes in #2969
• Pin unitxt to most recent minor version to avoid test failures by @kiersten-stokes in #2970
• mmlu pro generation_kwargs until Q: -> Question: by @yoonniverse in #2945
• AfroBench: How Good are Large Language Models on African Languages? by @JessicaOjo in #2825
• Added C4 Support by @Zephyr271828 in #2889
• Fixed a bug that in MMLU-Pro utils.py that throw index error if one choice was removed by @sleepingcat4 in #2870
• Add question suffix before the <|assistant|> tag by @TingchenFu in #2876
• Add device arg to model_args passed to LLM object in VLLM model class by @momentino in #2879
• paws-x fix formatting by @baberabb in #2759
• Delete scripts/cost_estimate.py by @StellaAthena in #2985
• Adding ACPBench Hard tasks by @harshakokel in #2980
• [SGLANG] Add the SGLANG generate API by @baberabb in #2997
• fix example notebook by @baberabb in #2998
• Log tokenized request warning only once by @RobGeada in #3002
• [Add Dataset Update] KBL 2025 by @abzb1 in #3000
• Output path fix by @Niccolo-Ajroldi in #2993
• use images with api models by @baberabb in #2981
• Adding resize images support by @artemorloff in #2958
• Revert "feat: add question suffix (#2876)" by @baberabb in #3007
• [hotfix] modify multimodal check in evaluate by @baberabb in #3013
• [Fix] Update resolve_hf_chat_template arguments by @fxmarty-amd in #2992
• Fix error due in Collating queries with different continuation lengths (fixes #2984) by @ameyagodbole in #2987
• [vllm] data parallel for V1 by @baberabb in #3011
• add arab_culture task by @bodasadallah in #3006
• chore: clean up and extend .gitignore rules by @e1washere in #3030
• Enable text-only evals for VLM models by @ysulsky in #2999
• [Fix] acc_mutual_info metric calculation bug by @baberabb in #3035
• Fix: fix vllm issue with DP>1 by @younesbelkada in #3025
• add Mbpp instruct by @baberabb in #2995
• remove prints by @baberabb in #3041
• [longbench] fix metric calculation by @baberabb in #2983
• Fallback to super implementation in fewshot_context for Unitxt tasks by @kiersten-stokes in #3023
• Fix Typo in README and Comment in utils_mcq.py by @vtjl10 in #3057
• fix longbech citation by @baberabb in #3061
• mmlu task: update README.md by @annafontanaa in #3070
• Fix typos in docstrings in instructions.py by @maximevtush in #3060
• bump version to 0.4.9 by @baberabb in #3073

New Contributors

• @wangcho2k made their first contribution in #2761
• @ruivieira made their first contribution in #2778
• @perlitz made their first contribution in #2773
• @kdymkiewicz made their first contribution in #2768
• @PabloAgustin made their first contribution in #2714
• @ksurya made their first contribution in #2727
• @zhuzeyuan made their first contribution in #2786
• @daniel-salib made their first contribution in #2785
• @oskarvanderwal made their first contribution in #1185
• @Avelina9X made their first contribution in #2802
• @agromanou made their first contribution in #2769
• @whoisjones made their first contribution in #2788
• @jd730 made their first contribution in #2813
• @anmarques made their first contribution in #2797
• @zhangruoxu made their first contribution in #2810
• @heli-qi made their first contribution in #2811
• @Tautorn made their first contribution in #2828
• @djwackey made their first contribution in #2854
• @Aprilistic made their first contribution in #2849
• @harshakokel made their first contribution in #2807
• @hadi-abdine made their first contribution in #2521
• @dazipe made their first contribution in #2824
• @danielholanda made their first contribution in #2866
• @Saibo-creator made their first contribution in #2865
• @houseroad made their first contribution in #2873
• @felipemaiapolo made their first contribution in #2520
• @dtrifiro made their first contribution in #2899
• @jerryzh168 made their first contribution in #2842
• @vmkhlv made their first contribution in #2919
• @annafontanaa made their first contribution in #2940
• @booxter made their first contribution in #2634
• @llsj14 made their first contribution in #2972
• @yoonniverse made their first contribution in #2945
• @Zephyr271828 made their first contribution in #2889
• @sleepingcat4 made their first contribution in #2870
• @TingchenFu made their first contribution in #2876
• @momentino made their first contribution in #2879
• @abzb1 made their first contribution in #3000
• @Niccolo-Ajroldi made their first contribution in #2993
• @fxmarty-amd made their first contribution in #2992
• @ameyagodbole made their first contribution in #2987
• @e1washere made their first contribution in #3030
• @ysulsky made their first contribution in #2999
• @younesbelkada made their first contribution in #3025
• @vtjl10 made their first contribution in #3057
• @maximevtush made their first contribution in #3060

Full Changelog: v0.4.8...v0.4.9