Skip to content
PMetal

pmetal distill

Knowledge distillation — transfer knowledge from a teacher model to a smaller student.

Distill knowledge from a larger teacher model into a smaller student model. Supports online, offline, and progressive methods.

Usage

Section titled “Usage”
Terminal window
pmetal distill \
  --teacher <TEACHER_MODEL> \
  --student <STUDENT_MODEL> \
  --dataset <DATASET> \
  [OPTIONS]

Examples

Section titled “Examples”
Terminal window
# Online distillation (live teacher inference)
pmetal distill \
  --teacher Qwen/Qwen3-4B \
  --student Qwen/Qwen3.5-0.8B-Base \
  --dataset train.jsonl


# Offline distillation (cached logits)
pmetal distill \
  --teacher Qwen/Qwen3-4B \
  --student Qwen/Qwen3.5-0.8B-Base \
  --dataset train.jsonl \
  --offline \
  --offline-cache ./output/distilled/teacher_logits

Methods

Section titled “Methods”
MethodDescription
OnlineLive teacher inference during training — highest quality, slowest
OfflinePre-cached logits with compression — faster, uses more disk
ProgressiveGradually increase distillation difficulty

Loss Functions

Section titled “Loss Functions”

Token-Level

Section titled “Token-Level”
  • KL Divergence — Standard distribution matching
  • Jensen-Shannon — Symmetric divergence
  • Soft Cross-Entropy — Temperature-scaled cross-entropy
  • TVD — Total variation distance
  • Hinge Ranking — Rank-based loss
  • Logistic Ranking — Logistic rank loss

Hidden State

Section titled “Hidden State”
  • MSE — Mean squared error on hidden states
  • Cosine — Cosine similarity matching
  • L1 — L1 distance

See Also

Section titled “See Also”