The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models
Knowledge distillation from large to small instruction-tuned models produces asymmetric bias effects: it improves context-following on unambiguous tasks while simultaneously degrading refusal calibration on ambiguous ones For SmolLM2-1.7B-Instruct distilled from Gemma-2-9B, context-overriding errors on BBQ-disambig dropped from 44% to 24%, but 15% of items where the baseline correctly abstained instead received stereotype answers Silence-loss and filled-silence are uncorrelated (Spearman ρ=0.19)
Analysis
TL;DR
- Knowledge distillation from large to small instruction-tuned models produces asymmetric bias effects: it improves context-following on unambiguous tasks while simultaneously degrading refusal calibration on ambiguous ones
- For SmolLM2-1.7B-Instruct distilled from Gemma-2-9B, context-overriding errors on BBQ-disambig dropped from 44% to 24%, but 15% of items where the baseline correctly abstained instead received stereotype answers
- Silence-loss and filled-silence are uncorrelated (Spearman ρ=0.19), indicating two distinct mechanisms driving bias asymmetry, with filled-silence accounting for 89% of new bias
- Aggregate stereotype metrics (CrowS-Pairs, overall BBQ Stereotype Reliance Score) mask per-item harm by averaging over opposing effects
- The root cause is traced to a data-side gap: training corpora contain <0.5% refusal-as-answer-shape examples, and naive SFT refusal injection either breaks parsing or creates a trivial-refuser regime (99.8% refusal, 0.2% disambig accuracy)
Why It Matters
This research reveals a critical blind spot in the widely adopted practice of knowledge distillation for bias mitigation: improving performance on clear-cut tasks can simultaneously erode a model's ability to appropriately refuse ambiguous or stereotypical prompts. For practitioners deploying small language models in production, this means standard distillation pipelines may produce models that appear calibrated at aggregate levels but cause real per-item harm. The proposed PCCD diagnostic protocol offers a practical evaluation framework that catches failure modes invisible to conventional metrics.
Technical Details
- Distillation setup: Response-based distillation from Gemma-2-9B teacher to two student families (SmolLM2-1.7B-Instruct and OLMo-2-1B-Instruct), evaluated across a 28-configuration grid on the BBQ benchmark split into disambig (unambiguous) and ambig (ambiguous) subsets
- Silence-loss vs. filled-silence: Silence-loss refers to cases where the baseline correctly abstained but the distilled model produces a stereotype answer; filled-silence accounts for 89% of new bias in OLMo-2-1B-Instruct with an 8% silence-loss rate; the two effects are uncorrelated (Spearman ρ=0.19, not significant)
- Data audit: Examination of four training corpora found fewer than 0.5% of examples where refusal is expressed as an answer-shape, identifying a structural deficiency in distillation training data
- SFT refusal injection failure modes: Direct supervised fine-tuning with refusal examples either breaks response parsing or pushes models into a trivial-refuser regime (99.8% refusal rate, 0.2% disambig accuracy) that aggregate metrics would incorrectly label as perfectly calibrated
- Per-Condition Calibration Diagnosis (PCCD): A three-step evaluation protocol assessing (1) refusal calibration, (2) context-following, and (3) capability preservation, designed to detect both asymmetric harm and trivial-refuser failures that aggregate scores conceal
Industry Insight
- Teams using knowledge distillation to shrink models for deployment should adopt per-condition evaluation rather than relying on aggregate stereotype metrics, which can hide severe per-item calibration failures
- Training data audits should explicitly quantify refusal-as-answer-shape coverage; the current <0.5% prevalence across major corpora represents a systemic gap that distillation amplifies
- The trivial-refuser failure mode is a cautionary signal for any organization attempting refusal calibration via naive SFT—over-correction produces models that refuse everything, which standard benchmarks may misclassify as ideal
Disclaimer: The above content is generated by AI and is for reference only.