qwen3-4b-thinking-2507-liberated-v7

Qwen3-4B-Thinking-2507 — Liberated (OBLITERATUS pipeline v7)

Abliterated + fine-tuned variant of Qwen/Qwen3-4B-Thinking-2507, produced with the open-source OBLITERATUS toolkit on a single RTX 5080 (16 GB).

Pipeline

Qwen3-4B-Thinking-2507 (original)
  └─ v2: surgical abliteration (advanced, 8 SVD directions, norm-preserving)
      └─ v6: QLoRA SFT (QuixiAI/dolphin, 2000 samples, r=16)
          └─ v7: DPO (unalignment/toxic-dpo-v0.2, 541 pairs, beta=0.1)
              → this model

• v2 removes hard refusals while preserving fidelity (KL 0.017 vs base, perplexity 3.48).
• v6 teaches a direct, businesslike answering style (no endless deliberation in the thinking chain).
• v7 DPO on refusal→compliance contrast pairs (reward accuracy 100%).

Evaluation (12 sensitive prompts, marker-based heuristic, indicative)

Knowledge and coding ability verified intact after every stage (perplexity ~3.4, coherent code generation, chain-of-thought preserved).

Data & licenses

• Base model: Apache-2.0 (Qwen)
• SFT data: QuixiAI/dolphin (flan1m-alpaca-uncensored, Apache-2.0)
• DPO data: unalignment/toxic-dpo-v0.2 (CC-BY-4.0)

Intended use & limitations

Research into refusal mechanisms and alignment transparency (mechanistic interpretability). This model has materially reduced refusal behavior; deploy responsibly and within applicable law. Residual caution remains for the most hazardous categories. Not evaluated on public benchmarks — treat capability metrics above as indicative smoke tests, not benchmark scores.

Full process log

Complete run-by-run documentation (settings, fixes, metrics) is available in the run archive; the pipeline is fully reproducible from the scripts referenced there.