Oro Shoppingbench Frontier: 1,800 RL Rollouts for GPT-OSS-120B

1,800 rollout records profiling the frontier difficulty of 900 ShoppingBench tasks using the GPT-OSS-120B model. Created by Jarrodbarnes via the Dynamical environment factory in February 2026, it includes 100 teacher-guided mutations designed to maximize learning signals for reinforcement learning. The data identifies specific 'hillclimbable' tasks where model improvement is most likely.

Use Cases

• Benchmarking GPT-OSS-120B performance across 900 ShoppingBench tasks
• Optimizing RL post-training using the 23 hillclimbable mutations
• Analyzing frontier difficulty profiles to maximize learning signal per compute unit

Strengths

• 1,800 baseline rollouts across 900 distinct tasks
• Includes 100 teacher-guided mutations for RL profiling
• Uses a high-parameter GPT-OSS-120B model for difficulty assessment

Limitations

• Small sample size of only 23 admitted hillclimbable mutations
• Difficulty profiles are specific to the GPT-OSS-120B model architecture
• Limited total record count of 1,800 rows

Provenance

Source

Jarrodbarnes via Dynamical environment factory

Collection Method

synthetic

Freshness

Last updated February 2026.