COMPREHENSIVE ANALYSIS: THE CALIBRATION PHASE (EXP01–07) I built an epistemic gate to stop LLM data poisoning during fine-tuning

说明与用途

作者构建了一个epistemic门，通过使用受限Softplus惩罚和基于向量的过滤器，阻止LLM微调过程中的数据中毒，防止损失崩溃并提升困惑度。

问题

Fine‑tuning large language models (LLMs) can be poisoned by false data, causing loss collapse and model degradation.

结果

The calibrated epistemic gate with Softplus penalty and vector similarity prevented loss collapse, achieved stable margins (~+10), and reduced perplexity when combined with LoRA.

工作流

• 1. Define loss functions and filtering mechanisms (EXP01–04).
• 2. Test raw negative log‑prob loss leading to collapse (EXP03).
• 3. Replace with bounded Softplus penalty (EXP05).
• 4. Transition from dictionary lookup to vector gate using cosine similarity (EXP07).
• 5. Validate with real‑world facts and reproducibility checks (EXP06–07).
• 6. Compare architectures (LoRA vs. π_ref) to address perplexity (EXP08–09).

证据

The loss diverges to −27, margin reaches +13, but the model destroys itself—the unbounded negative gradient is a bomb

CE(truth) + β•softplus(m + logP(false) − logP(truth)) ✨ The synthesis: relative, bounded punishment. It never changed again

EXP07: Embeddings + cosine similarity (DenseVectorGate)—and the margins suddenly stabilize: +10.2/+10.1/+10.6