SWE-bench Verified: Trajectory Embeddings from Qwen3.6-27B for 10 Models

Pre-computed pooled embeddings of mini-swe-agent trajectories on the SWE-bench Verified benchmark. The embeddings were generated using the Qwen/Qwen3.6-27B model in a vLLM runner with last-token pooling. The dataset was authored by tarsur385 and last updated on 2026-05-18.

Use Cases

• Analyzing the performance of different code-generation models based on their trajectory embeddings.
• Training or fine-tuning models for software engineering tasks using pre-computed trajectory features.
• Comparing the behavioral patterns of AI agents on the SWE-bench Verified benchmark.
• Developing similarity metrics or clustering algorithms for agent trajectories in code problem-solving.

Strengths

• Embeddings are derived from a specific, high-capacity model (Qwen/Qwen3.6-27B).
• Trajectories are based on the established SWE-bench Verified benchmark for software engineering.
• The dataset includes embeddings for 10 different models, allowing for comparative analysis.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Pre-computed from mini-swe-agent trajectories using Qwen/Qwen3.6-27B in vLLM runner with last-token pooling.

Freshness

Last updated 2026-05-18 07:38:49; freshness should be verified.