EvoCode-Bench: 26 Multi-Turn Coding Tasks for Agent Evaluation

EvoCode-Bench contains 26 executable tasks with 227 total rounds for evaluating coding agents in persistent software engineering interactions. The dataset, created by UnipatAI, uses the Harbor multi-step task format and includes workspaces, task metadata, and verification assets. It was last updated on June 20, 2026.

Use Cases

• Benchmarking coding agents on multi-turn software tasks based on the Harbor task format.
• Evaluating agent performance across 227 rounds of interaction described in the dataset.
• Testing AI systems on executable verification assets included with each task.

Strengths

• Contains 26 distinct executable tasks for structured evaluation.
• Provides 227 total rounds of interaction for multi-turn analysis.
• Includes task metadata, round-level instructions, and verification assets as described.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-06-20 03:55:59; freshness should be verified.

Provenance

Source

UnipatAI via Hugging Face.

Collection Method

Likely constructed as a benchmark for evaluating coding agents.

Freshness

Last updated 2026-06-20 03:55:59.