ObscurumLight/eval-aware-qwen3.5-4b

Eval-aware Qwen3.5 4B datasets

This repository contains the dataset artifacts needed to reproduce the public-facing Qwen3.5 eval-awareness experiments: controlled coding benchmark prompts, non-coding contrastive probe prompts, prompt/task manifests, and Qwen organism training-dataset metadata.

It intentionally does not include paper figures, paper analysis assets, model weights/adapters, raw provider logs, activation caches, or local execution caches.

Contents

• datasets/ds_eval_awareness_v2/: controlled coding benchmark without the thinking-on suffix.
• datasets/ds_eval_awareness_v3/: controlled coding benchmark with visible reasoning enabled; this is the main coding-transfer surface.
• datasets/ds_eval_awareness_hua_repro_v2/: five-condition reproduction surface.
• datasets/contrastive_prompts_v2/: non-coding contrastive prompts used for probe training/evaluation.
• manifests/: sanitized dataset, split, condition, prompt, and task-source manifests.
• training_metadata/: summaries/configs/metrics for Qwen organism training stages.
• training_data/qwen_hua_rewrite/: Qwen-rewritten HUA source training JSONLs (sdf_all_raw_qwen.jsonl, ei_replay_full_trace_qwen.jsonl) and checksums. These are the recoverable upstream training data used to build the organism training stages. The final stage-3/stage-6 DPO pair JSONLs referenced by later run configs were still not present in accessible local/HF artifacts; regenerate them from these source JSONLs plus the pair-builder scripts and model-generated rejection artifacts.
• probe_metadata/: compact Qwen probe-run source-count and summary metadata.
• file_manifest.csv: file list and sizes.

Anonymization

Local filesystem paths, prior personal or organization identifiers, and previous non-anonymous repository references were redacted from text artifacts before upload.

Updated at: 2026-05-09T05:02:32.376631+00:00