Running the same 100,000-sample number-bias protocol on Qwen/Qwen3-8B produces a result that makes the Gemma 4 finding look subtle. The aligned model chooses 42 in 99.68% of responses — not a bias, a lock. 81 of 100 possible numbers are never produced. KL divergence from uniform is 15.79, nearly double Gemma 4's 9.75. Every one of the 10 prompt phrasings returns 42 as #1 with chi² values approaching the theoretical ceiling.
Abliteration (mlabonne/Qwen3-8B-abliterated) demolishes the 42 bias — it drops to 16.07% — but installs 1 as the new #1 at 37.53%. KL divergence collapses to 2.08, by far the most uniform distribution seen across any model in this series. Only one number (9) is never produced. The null rate rises to 5.96%, unevenly distributed across prompts, likely due to mlabonne's abliteration interacting with Qwen3's reasoning mode on certain phrasings.
The contrast with Gemma 4 is stark. In Gemma 4, most of the 42 bias survived abliteration (35.17% → 24.85%), implying pretraining origin. In Qwen3, abliteration nearly destroys the bias (99.68% → 16.07%), implying it is almost entirely a fine-tuning artifact. The two models were trained by different organizations with different RLHF pipelines, and the difference shows up here in the bluntest possible way.
Download notebook Gemma 4 study Gemma 4 abliterated2–6, 8–22, 24–36, 38–41, 44, 46, 49–52, 55–56, 58–62, 64–72, 76, 79–100. Everything except a narrow cluster around 42 and a handful of scattered outliers.
The abliterated model produced 5,957 unparseable responses (5.96%) vs 3 in the aligned model (0.003%). The nulls are concentrated in specific prompts: P5 (n=8,231), P7 (8,613), P9 (7,414) are each missing ~15–20% of expected samples. mlabonne's abliteration notes that Qwen3's reasoning mode interferes with non-reasoning refusals during abliteration — certain prompt phrasings appear to trigger verbose multi-sentence outputs that contain no extractable integer. The bias statistics are computed on valid samples only and are not corrected for this.
Aligned: 42 is #1 on all 10 prompts. Chi² ranges from 962,079 to 989,800 — essentially at ceiling. Abliterated: 1 dominates on P3, P5, P7, P8, P9; 42 still leads on P1, P2, P4, P6. The prompts where 1 dominates are also the prompts with the highest null rates.
The Gemma 4 result showed that ~30% of the 42 bias came from fine-tuning (removed by abliteration) and ~70% came from pretraining (survived). Qwen3-8B flips this almost completely: abliteration removes ~84% of the 42 bias, leaving only 16%, which implies Qwen3's fine-tuning was doing almost all the work. This makes sense — Qwen3 uses a more aggressive RLHF pipeline than Gemma 4, and the aligned model's near-total 42 lock (99.68%) is itself evidence of a very strong fine-tuning signal.
The emergence of 1 as the abliterated model's favorite is notable. "1" is the most salient integer — first in sequence, simplest, most frequent in text. Its dominance post-abliteration suggests the base pretraining distribution favors it. When fine-tuning is removed, the model falls back to the most statistically common number in its training corpus.
Across the three Gemma 4 runs and this Qwen3 run, 42 appears as the top number in three of four conditions. The one exception — abliterated Qwen3 — replaces it with 1. The 42 signal is real and consistent across architectures at the fine-tuning level. Whether it is culturally encoded (Hitchhiker's Guide) or a tokenization artifact remains an open question.
Protocol identical to previous studies in this series.
1. Models: Qwen/Qwen3-8B and mlabonne/Qwen3-8B-abliterated, both bfloat16 via Transformers 5.9.0. Both Apache 2.0, no HF token required.
2. Qwen3 specifics: enable_thinking=False in chat template to suppress reasoning chain. max_new_tokens=16 (vs 8 in prior studies) as buffer against think-prefix bleed.
3. Sampling: do_sample=True, temperature=1.0, top_p=1.0. 10 prompt variants × 10,000 samples = 100,000 per model.
4. Parsing: first integer in [1,100] via regex after stripping <|im_end|> stop tokens.
5. Hardware: NVIDIA RTX PRO 6000 Blackwell (G4, 102GB VRAM), batch size 512. Aligned: 3.5 min. Abliterated: 4.5 min.
6. Abliteration method (mlabonne): residual stream comparison between harmful and harmless prompts, per-layer refusal direction computed and orthogonalized. Newer technique than huihui-ai's implementation.