OpenThoughts-Agent-v1-RL: 720 Tasks with Verifiers for Agentic RL

OpenThoughts-Agent-v1-RL provides approximately 720 curated reinforcement learning tasks designed for training agentic models, released by the open-thoughts project in January 2026. The collection includes instructions, environment configurations, and verifiers specifically optimized for benchmarks like Terminal-Bench 2.0 and SWE-Bench.

Use Cases

• Training agents to execute terminal commands using the environment and verifier components
• Fine-tuning models for software engineering automation based on SWE-Bench style instructions
• Developing reward functions for RLHF using the success signals from the verifier fields

Strengths

• Contains ~720 curated tasks
• Includes integrated verifiers for automated performance evaluation
• Optimized for Terminal-Bench 2.0 and SWE-Bench standards

Limitations

• Small sample size of approximately 720 records
• Limited to text-based agentic modalities

Provenance

Source

open-thoughts project

Collection Method

curated

Freshness

Last updated January 2026.