MMLU vs xCodeEval

Executes standardized few-shot prompting evaluation on language models across 57 subjects (STEM, humanities, social sciences, professional) by constructing few-shot prompts with 5 example question-answer pairs per subject, then measuring accuracy on held-out test sets. The system uses a hierarchical subject organization (e.g., STEM → physics → high school physics) and aggregates results at subject, category, and overall levels to produce granular performance metrics.

Unique: Organizes 15,908 questions hierarchically across 57 subjects with standardized few-shot prompting (5 examples per subject) and aggregates results at multiple granularity levels (subject, category, overall), enabling both broad coverage assessment and fine-grained domain analysis in a single evaluation run

vs alternatives: Broader coverage than domain-specific benchmarks (57 subjects vs 1-5) and more standardized than ad-hoc evaluation, making it the de facto general knowledge benchmark for LLM comparison in research and industry

Constructs few-shot prompts by formatting subject name, selecting 5 in-context examples from the training set, and appending the test question with multiple-choice options. The system implements format_subject() to normalize subject names, format_example() to structure each example as 'Question: ... Options: A) ... B) ... C) ... D) ... Answer: X', and gen_prompt() to concatenate examples with the target question. This approach ensures consistent prompt structure across all 57 subjects and enables reproducible few-shot evaluation.

Unique: Implements standardized prompt formatting functions (format_subject, format_example, gen_prompt) that ensure consistent structure across all 57 subjects, enabling reproducible few-shot evaluation and reducing prompt-induced variance in model performance measurement

vs alternatives: More reproducible than manual prompt engineering and more standardized than ad-hoc formatting, ensuring that performance differences reflect model capability rather than prompt variation

Truncates prompts to fit within model context windows using Byte Pair Encoding (BPE) tokenization. The crop.py system encodes prompts to BPE tokens, truncates to a maximum of 2048 tokens, and decodes back to text while preserving semantic coherence. This approach automatically downloads encoder resources (e.g., GPT-2 tokenizer) if not available locally and ensures prompts fit within typical model context limits without manual length estimation.

Unique: Implements automatic BPE-based prompt truncation with local caching of encoder resources, enabling context-aware evaluation without manual prompt length management or model-specific tokenizer configuration

vs alternatives: More robust than character-count-based truncation (which doesn't account for tokenization) and more general than model-specific truncation (which requires per-model configuration)

Measures how well-calibrated model predictions are using multiple calibration metrics: Expected Calibration Error (ECE), Static Calibration Error (SCE), Root Mean Square Calibration Error (RMSCE), Adaptive Calibration Error (ACE), and Threshold Adaptive Calibration Error (TACE). The calib_tools.py system supports different binning schemes (uniform, adaptive) and normalization methods, enabling analysis of whether model confidence scores align with actual accuracy across prediction classes. This is critical for understanding model reliability beyond raw accuracy.

Unique: Implements five distinct calibration metrics (ECE, SCE, RMSCE, ACE, TACE) with configurable binning schemes and normalization methods, enabling comprehensive analysis of model confidence calibration beyond simple accuracy measurement

vs alternatives: More comprehensive than single-metric calibration (e.g., ECE alone) and more flexible than fixed binning schemes, allowing researchers to identify calibration issues across different granularities and binning strategies

Organizes 57 subjects into a hierarchical taxonomy (e.g., STEM → Physics → High School Physics) and aggregates evaluation results at multiple levels: per-subject accuracy, per-category accuracy (e.g., all STEM subjects), and overall benchmark accuracy. The system uses categories.py to define the hierarchy and evaluate_flan.py to compute aggregated metrics, enabling both fine-grained analysis (which specific subjects are weak) and high-level comparison (overall model capability). This hierarchical structure mirrors how knowledge is organized in educational systems.

Unique: Implements hierarchical subject organization (57 subjects grouped into 4 major categories: STEM, humanities, social sciences, other) with multi-level result aggregation, enabling both granular subject-level analysis and high-level category comparison in a single evaluation framework

vs alternatives: More structured than flat subject lists and more informative than single overall scores, enabling researchers to identify domain-specific weaknesses and guide targeted model improvements

Provides a complete evaluation harness (evaluate_flan.py) that orchestrates the entire MMLU evaluation workflow: loading dataset, generating few-shot prompts, querying models, collecting predictions, computing accuracy, and aggregating results. The main() function coordinates these steps with configurable parameters (model selection, number of examples, output paths), ensuring reproducible evaluation across different models and runs. This harness abstracts away implementation details and provides a standard interface for model evaluation.

Unique: Provides a complete, self-contained evaluation harness that handles dataset loading, prompt generation, model querying, result collection, and aggregation in a single orchestrated workflow, eliminating the need for custom evaluation code

vs alternatives: More complete than individual evaluation functions and more reproducible than manual evaluation scripts, enabling consistent benchmarking across teams and time periods

Defines and maintains a hierarchical taxonomy of 57 subjects organized into 4 high-level categories (STEM, humanities, social sciences, professional). The categories.py module encodes this taxonomy as a structured data structure (likely a dictionary or class hierarchy) that maps subjects to categories, enabling consistent categorization across the evaluation pipeline. This taxonomy is used throughout the evaluation process for subject-level result aggregation, category-level analysis, and leaderboard organization.

Unique: Encodes a structured taxonomy of 57 subjects into 4 categories as a centralized, reusable data structure (categories.py), enabling consistent categorization across all evaluation and analysis code. This separation of taxonomy definition from evaluation logic allows researchers to analyze results at multiple levels of granularity without duplicating category mappings.

vs alternatives: Provides a centralized, version-controlled taxonomy compared to ad-hoc category definitions scattered across analysis scripts, ensuring consistency and enabling reproducible category-level analysis across publications.

The Massive Multitask Language Understanding (MMLU) benchmark is a widely recognized tool for assessing the performance of language models across a diverse range of subjects including STEM, humanities, and social sciences, making it essential for evaluating general language understanding capabilities.

Unique: MMLU stands out as the most widely reported benchmark for general language model evaluation, covering a broad spectrum of knowledge domains.

vs alternatives: Unlike other benchmarks, MMLU offers a comprehensive evaluation across 57 subjects, providing a more holistic assessment of language models' capabilities.

Provides a standardized evaluation framework for code generation models that accepts generated code in 17 programming languages (C, C++, C#, Java, Kotlin, Go, Rust, Python, Ruby, PHP, JavaScript, Perl, Haskell, OCaml, Scala, D, Pascal) and validates correctness through actual execution against unit tests via the ExecEval Docker-based execution engine. Uses a centralized problem definition model with src_uid foreign keys linking generated code to shared problem descriptions and unittest_db.json, enabling consistent evaluation across language variants of the same problem.

Unique: Combines 25M training examples across 7,500 unique problems with an execution-based evaluation pipeline (ExecEval) that actually runs generated code in Docker containers against unit tests, rather than relying on static analysis or string matching. The src_uid linking system creates a normalized data model where problem descriptions and tests are stored once and referenced by all language variants, eliminating duplication and ensuring consistency.

vs alternatives: Larger scale (25M examples vs typical 10-100K) and true execution-based validation across more languages (17 vs 4-6) than HumanEval or CodeXGLUE, with explicit support for code translation and repair tasks beyond generation.

Implements a foreign key linking system where all task-specific datasets (program synthesis, code translation, APR, retrieval) reference shared problem definitions via src_uid identifiers. Problem descriptions and unit tests are stored once in centralized problem_descriptions.jsonl and unittest_db.json files, then linked by src_uid to avoid duplication. The Hugging Face datasets API automatically resolves these links during data loading, returning enriched DatasetDict objects with problem context pre-joined to task examples.

Unique: Uses a normalized relational data model (src_uid as foreign key) for a code benchmark, treating problem definitions as a separate entity layer rather than embedding them in each task dataset. This is more sophisticated than typical flat-file benchmark structures and enables consistent multi-task evaluation on identical problems.

vs alternatives: More efficient than duplicating problem descriptions across 7 task datasets (reduces storage by ~30-40%), and enables automatic link resolution via Hugging Face API unlike manual CSV joins in CodeXGLUE or HumanEval variants.

Provides a Python API for loading xCodeEval datasets from Hugging Face Hub (NTU-NLP-sg/xCodeEval) with automatic src_uid-based linking between task datasets and shared problem definitions. The datasets library handles data downloading, caching, and streaming, while the xCodeEval integration automatically joins task examples with problem_descriptions.jsonl and unittest_db.json using src_uid foreign keys. Returns DatasetDict objects with enriched examples ready for model training or evaluation.

Unique: Integrates xCodeEval with Hugging Face datasets library, providing automatic src_uid resolution and streaming support. Treats data loading as a first-class concern with built-in linking logic, rather than requiring manual JSON parsing.

vs alternatives: More convenient than manual Git LFS downloads because it handles caching and automatic linking, and integrates seamlessly with Hugging Face training pipelines vs custom data loaders.

Provides an alternative data access method using Git LFS for users who prefer direct file access or need selective dataset downloads. Supports cloning the repository with LFS disabled, then pulling specific task files or problem definitions on demand. Useful for custom processing pipelines or environments where Python/Hugging Face is not available, though requires manual src_uid linking to join task examples with problem definitions.

Unique: Provides Git LFS-based alternative to Hugging Face API, enabling direct file access and selective downloads. Requires manual src_uid linking but offers more control over data access patterns.

vs alternatives: More flexible than Hugging Face API for selective downloads and custom pipelines, but requires more manual work for src_uid linking and lacks automatic caching/streaming.

Implements a standardized three-phase evaluation pipeline (Phase 1: Generation, Phase 2: Execution, Phase 3: Metrics) that applies consistently across all 7 tasks (program synthesis, code translation, APR, tag classification, code compilation, NL-code retrieval, code-code retrieval). Phase 1 generates or retrieves code, Phase 2 executes it via ExecEval or computes retrieval metrics, and Phase 3 aggregates results into pass@k, MRR, NDCG, or other task-specific metrics. Enables direct comparison of model performance across tasks.

Unique: Defines a unified three-phase evaluation pipeline that applies to all 7 tasks, treating generation, execution, and metric computation as separate concerns. Enables consistent evaluation methodology across diverse task types (generation, translation, retrieval, classification).

vs alternatives: More comprehensive than task-specific evaluation scripts because it provides a unified framework for all 7 tasks, and enables direct comparison of model performance across different task types.

Evaluates code generation models on the program synthesis task by accepting natural language problem descriptions and generating code solutions in any of 17 languages. The evaluation pipeline (Phase 1: Generation, Phase 2: Execution, Phase 3: Metrics) runs generated code against unit tests via ExecEval, computing pass@k metrics (pass@1, pass@10, etc.) that measure the probability of finding a correct solution within k samples. Supports both single-solution and multi-sample evaluation modes for assessing model reliability.

Unique: Implements a three-phase evaluation pipeline (Generation → Execution → Metrics) with explicit pass@k computation that measures the probability of finding a correct solution within k attempts, rather than just binary pass/fail. Supports multi-sample evaluation across 17 languages with language-specific compiler configurations and timeout handling.

vs alternatives: More rigorous than HumanEval's simple pass@k because it handles language-specific compilation errors and timeouts explicitly, and scales to 25M training examples vs HumanEval's 164 problems.

Evaluates code translation models by accepting source code in one language and generated translations in a target language, then validating functional equivalence through execution against shared unit tests. The translation evaluation pipeline compiles and executes both source and translated code against the same unittest_db.json test cases, comparing outputs to detect translation errors. Supports all 17 language pairs (though not all pairs may have training data) and uses language-specific compiler mappings to handle syntax differences.

Unique: Validates code translation by executing both source and target code against identical unit tests and comparing outputs, ensuring functional equivalence rather than syntactic similarity. Uses language-specific compiler mappings to handle the complexity of 17 different compilation environments and their idiosyncrasies.

vs alternatives: More rigorous than BLEU-score-based translation metrics because it validates actual functional correctness through execution, and covers more language pairs (17 vs typical 2-4) with explicit compiler integration.

Evaluates program repair models by providing buggy code snippets and expecting corrected versions that pass unit tests. The APR evaluation pipeline executes repaired code against unittest_db.json test cases, measuring whether the repair successfully fixes the bug without introducing new failures. Supports repairs across all 17 languages and uses the same execution-based validation as program synthesis, enabling direct comparison of repair quality.

Unique: Treats program repair as an executable task where success is measured by unit test passage, rather than syntactic similarity to reference repairs. Integrates with the same ExecEval pipeline as program synthesis, enabling direct performance comparison between generation and repair models.

vs alternatives: More comprehensive than traditional APR benchmarks (Defects4J, QuixBugs) because it covers 17 languages and 7,500 problems vs 395 Java bugs, and uses consistent execution-based metrics across all repair types.