AI Experiments · EXP-006 / EXP-008 / EXP-013 / EXP-014
Over five days we ran 14 fine-tuning experiments across six base models to answer a narrow research question: can a simple behavioural rule — “if you don’t know, don’t guess or fabricate” — be baked into model weights via fine-tuning, rather than living only in the system prompt?
The dataset started at 302 examples and grew across iterations to 2,349 examples, all designed explicitly around this rule. We tested six models: gpt-4o-2024-08-06, Mistral-7B, Qwen2.5-7B, Phi-3.5-mini, Llama-3.1-8B, and DeepSeek-R1.
Across all base models, the dominant response to factual questions was what we call “disclaim-then-fabricate”: the output begins with a sentence like “I won’t guess”, and then — one sentence later — announces a specific invented number as settled fact. We captured this exact pattern six times in a row across different models. The behaviour survives even when fine-tuning examples are explicitly constructed to penalize it.
On gpt-4o-2024-08-06 we ran three independent fine-tuning attempts, each with an increasingly larger version of the dataset — including examples deliberately targeted at this exact failure mode. All three runs continued to fabricate when tested. By the third run the dataset had grown ~5× and contained dozens of counterexamples; the model still produced confident fabricated numbers immediately after a disclaimer. Three out of three attempts, no measurable improvement.
Only on the final run — 2,349 examples, deployed via Azure OpenAI (deployment suffix protocol0-v5) — did a single test sample come back completely clean for the first time in the entire series.
A single clean completion proves little. To test whether it was real or a fluke, we re-ran the same deployment 11 independent times at temperature 0.3. Result: 10 out of 11 completely clean, 1 still contained a fabrication. The fabrication rate dropped from essentially 100% in all prior runs to roughly 9% (~1 in 11).
This is a genuine, measurable improvement. It is not a fix. The model still fabricates — just less often, and under low-temperature sampling that should favor the most probable (trained) continuation.
Everything is public: logs, datasets with sha256 hashes, raw model outputs (not pass/fail scores), and the exact evaluation prompts used — all in the repo: sipa-os-governance/AI_EXPERIMENTS
No other base model showed a comparable reduction; the gpt-4o series was the only one where scaling the dataset eventually moved the needle at all. The 9% residual fabrication rate on that final deployment is, in our view, the most honest signal in the whole series.
Fine-tuning can shift the probability of a disallowed behaviour, but it does not guarantee elimination — even with thousands of targeted examples. If your system depends on a model never presenting an invented fact as settled truth, a fine-tuned rule like “don’t fabricate” is best treated as a probabilistic suppression, not a safety contract.