Research map / Image generation / Autoregressive and latent generation
EpiCoder: Encompassing Diversity and Complexity in Code Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a feature tree-based synthesis framework that extracts hierarchical code features from The Stack v2, evolves them in depth and breadth, and samples subtrees to generate 433k instruction samples (function- and file-level) for fine-tuning EpiCoder models.
Problem
Code instruction-data synthesis from code snippet seeds limits the complexity and diversity of generated data, and existing methods rarely produce multi-file, repository-level tasks.
Approach
Organize code features into a hierarchical feature tree extracted by an LLM from raw code, iteratively expand it in breadth and depth, reweight sampling probabilities with a temperature, and sample subtrees of controllable depth/breadth to drive task and code generation (including multi-file solutions).
Key results
- EpiCoder-Qwen-7B reaches 59.0 average pass@1 across function-level benchmarks, the highest among compared ~7B models (Qwen2.5-Coder-7B-Instruct: 57.0).
- EpiCoder-Qwen-7B scores 27.7 on BigCodeBench-Hard Complete, about 7.4 points above Qwen2.5-Coder-7B-Instruct.
- EpiCoder-DS-6.7B-75k averages 51.2 versus 48.2 for WaveCoder-Ultra-6.7B (130k) on function-level benchmarks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune Qwen2.5-Coder-7B on EpiCoder synthesized instruction data | Qwen2.5-Coder-7B · Function-level benchmarks, average | Pass@1 (%): 57.0 -> 59.0 | Helped |
| Fine-tune DeepSeek-Coder-Base-6.7B on EpiCoder synthesized instruction data | DeepSeek-Coder-Base-6.7B · Function-level benchmarks, average | Pass@1 (%): 48.1 -> 53.1 | Helped |
| Use 75k EpiCoder function-level samples instead of Magicoder-DS (75k) | DeepSeek-Coder-Base-6.7B · Function-level benchmarks, average | Pass@1 (%): 45.8 -> 51.2 | Helped |
| Use 75k EpiCoder function-level samples instead of WaveCoder-Ultra-6.7B data | DeepSeek-Coder-Base-6.7B · Function-level benchmarks, average | Pass@1 (%): 48.2 -> 51.2 | Helped |
| Use 74k EpiCoder function-level samples instead of SelfCodeAlign data | CodeQwen-7B · Function-level benchmarks, average | Pass@1 (%): 58.6 -> 62.6 | Helped |
| Train on EpiCoder file-level data | EpiCoder · XFileDep | Pass@1 | Helped |
| Generate function-level code with the feature tree synthesis framework rather than existing code datasets | EpiCoder function-level data · Code complexity (LLM judge) | Average score: 32.6% | Helped |
| Generate file-level code with the feature tree synthesis framework rather than existing code datasets | EpiCoder file-level data · Code complexity (LLM judge) | Average score: 52.5% | Helped |
Surprising findings
- Function-level data scores higher than existing datasets on Halstead complexity, but the gains are largest at file level, suggesting multi-file data drives complexity.
Limitations
- Relies on GPT-4o for feature extraction, evolution, and generation, so synthesis cost and dependence on a proprietary model are not quantified.
- Repository-level generation is shown only qualitatively via one example repository, with no quantitative evaluation.
- Complexity and diversity are assessed by GPT-4o-as-judge and feature counts extracted by GPT-4o, which may be biased.
Experimental setup
Datasets: The Stack v2 evol-codealpaca-v1
Benchmarks: HumanEval MBPP BigCodeBench EvoEval FullStackBench XFileDep
Metrics: pass@1
Baselines: Qwen2.5-Coder-7B-Instruct DeepSeekCoder-6.7B-Instruct Magicoder-S-DS WaveCoder-Ultra-6.7B OpenCodeInterpreter-DS-6.7B OpenCoder-8B-Instruct SelfCodeAlign-CQ-7B Magicoder-DS
Builds on: Magicoder Evol-Instruct OSS-Instruct SelfCodeAlign
Closest papers by meaning
- Divide-and-Conquer Meets Consensus: Unleashing the Power of Functions in Code Generation (NeurIPS 2024 · similarity 0.85)
- Magicoder: Empowering Code Generation with OSS-Instruct (ICML 2024 · similarity 0.85)
- Let's Revise Step-by-Step: A Unified Local Search Framework for Code Generation with LLMs (NeurIPS 2025 · similarity 0.84)
- CodeChemist: Test-Time Scaling for Low-Resource Code Generation via Functional Knowledge Transfer (ICML 2026 · similarity 0.84)
- Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks (NeurIPS 2025 · similarity 0.84)
- Repoformer: Selective Retrieval for Repository-Level Code Completion (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.