The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models

知识蒸馏对小型语言模型偏见的不对称影响

Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias. On unambiguous tasks (BBQ-disambig), response-based distillation from a Gemma-2-9B teacher improves context-following: for the most biased baseline (SmolLM2-1.7B-Instruct), it cuts the context-overriding error rate from 44% to 24%.

摘要： 我们研究发现，在小型指令微调语言模型中，知识蒸馏对偏见具有不对称的影响。在无歧义任务（BBQ-disambig）中，基于 Gemma-2-9B 教师模型的响应式蒸馏提升了模型对上下文的遵循能力：对于偏见最严重的基准模型（SmolLM2-1.7B-Instruct），它将上下文覆盖错误率从 44% 降低到了 24%。

On ambiguous tasks (BBQ-ambig), the same distillation destroys per-item refusal calibration: 15% of items where the baseline correctly abstained instead receive stereotype answers, even when overall refusal rate is preserved. The pattern reproduces on a second student family (OLMo-2-1B-Instruct), with silence-loss of 8% and filled-silence accounting for 89% of new bias.

在有歧义任务（BBQ-ambig）中，同样的蒸馏过程却破坏了单项拒绝校准：在基准模型原本能正确拒绝回答的样本中，有 15% 的样本转而输出了刻板印象答案，即便整体的拒绝率保持不变。这种模式在第二个学生模型系列（OLMo-2-1B-Instruct）中也得到了重现，其中“沉默丢失”（silence-loss）为 8%，而“填充式沉默”（filled-silence）占新增偏见的 89%。

Across the full 28-configuration grid, the magnitudes of silence-loss and filled-silence are uncorrelated (Spearman $\rho=0.19$, n.s.), indicating that the two effects arise from distinct mechanisms. Aggregate stereotype metrics (CrowS-Pairs, overall BBQ Stereotype Reliance Score) average over both effects and conceal the per-item harm.

在全部 28 种配置网格中，“沉默丢失”和“填充式沉默”的量级并不相关（Spearman $\rho=0.19$, n.s.），这表明这两种效应源于不同的机制。聚合的刻板印象指标（如 CrowS-Pairs、整体 BBQ 刻板印象依赖得分）对这两种效应进行了平均化处理，从而掩盖了单项样本中的危害。

We trace the calibration loss to a data-side mechanism: an audit of four training corpora finds <0.5% refusal-as-answer-shape. Supervised fine-tuning (SFT) with refusal injection either breaks parsing or over-corrects into a trivial-refuser regime (refusal rate 99.8%, disambig accuracy 0.2%) that aggregate metrics would call perfectly calibrated.

我们将这种校准损失追溯到数据层面的机制：对四个训练语料库的审计发现，以“拒绝回答”作为答案形式的数据占比不到 0.5%。通过注入拒绝样本进行监督微调（SFT），要么会导致解析崩溃，要么会过度修正进入一种“无脑拒绝”状态（拒绝率 99.8%，无歧义任务准确率 0.2%），而聚合指标却会将其判定为完美校准。

We propose Per-Condition Calibration Diagnosis (PCCD), a three-step protocol that evaluates refusal calibration, context-following, and capability preservation. PCCD catches both the asymmetric harm and the trivial-refuser failure mode that aggregate evaluations miss.

我们提出了“分条件校准诊断”（PCCD），这是一种包含三个步骤的协议，用于评估拒绝校准、上下文遵循和能力保持。PCCD 能够捕捉到聚合评估所遗漏的不对称危害以及“无脑拒绝”的失效模式。