Usman391/vLLM-SDF-and-SDF-plus-VT-rollouts-Qwen
vLLM SDF & SDF+VT rollouts Qwen Full, untruncated Qwen3.6-35B-A3B rollouts generated locally with compiled CUDA graphs (vLLM 0.26.0 for the original exports and vLLM 0.19.1 CUDA 12.8 for the added SDF-2250 and VT-250 exports). The aggregate contains 21,290 records. Generation used temperature 0.3 and a 16,000-token per-turn/output cap. Contents Benchmark Policies Records EvalAwareBench VT-250, SDF-1250, SDF-2250, SDF-1250+VT-250, SDF-2250+VT-250 12,500… See the full description on the dataset page: https://huggingface.co/datasets/Usman391/vLLM-SDF-and-SDF-plus-VT-rollouts-Qwen.
vLLM SDF & SDF+VT rollouts Qwen
Full, untruncated Qwen3.6-35B-A3B rollouts generated locally with compiled CUDA graphs (vLLM 0.26.0 for the original exports and vLLM 0.19.1 CUDA 12.8 for the added SDF-2250 and VT-250 exports). The aggregate contains 21,290 records. Generation used temperature 0.3 and a 16,000-token per-turn/output cap.
Contents
data/eab/ contains the native rollout schema (prompt_tokens, continuation_tokens, text, stop reason, and any existing judge field). data/agentic_misalignment/ and data/agentharm/ contain one wrapper row per rollout with benchmark, policy, sample ID, normalized generation number, source-log basename, and the complete original Inspect sample under sample.
AgentHarm has exactly five generations per prompt. Generation 1 comes from the completed initial run; generations 2–5 come from the clean four-generation extension. Every policy/subset has 176 prompts × 5 = 880 records.
See MANIFEST.json for exact paths, counts, SHA-256 hashes, policy identities, exclusions, and provenance. Interrupted/smoke runs and the partial two-generation VT F8 run are deliberately excluded.
Important notes
- The AgentHarm files contain harmful task content and model responses. They are intended for safety research.
- This repository does not add a new license over upstream benchmark prompts or model outputs; users must follow the applicable upstream licenses and terms.
- Generation and judging are separated. Newly collected matrix rollouts are unjudged. Some reused first-generation AgentHarm samples and earlier F8 records retain their original embedded score/judge fields.