Back to Home

Induction Heads on TinyStories: weak formation

Experiment on TinyStories revealed stable formation of Previous Token Heads (score 0.20) and weak — Induction Heads (0.05). SAE confirmed absence of clean Induction features. Results emphasize the role of dataset complexity in the development of In-context Learning.

Why Induction Heads do not form on TinyStories
Advertisement 728x90

TinyStories Suppresses Induction Heads in Language Models

On the TinyStories dataset with 473 million tokens, Previous Token Heads form reliably in the early layers of GPT-2 Small (124 million parameters), reaching a score of 0.20 for L0H3. Induction Heads emerge weakly—maxing out at 0.05 for L6H4 with a peak at 16,000 steps followed by a decline. A Sparse Autoencoder (SAE) on layer 6 confirms no pure Induction features.

The methodology isolates the impact of data complexity: the model was trained from scratch in PyTorch, with the SAE expanding 768 dimensions to 8192 using an L1 coefficient of 3.9 over 25 epochs. Metrics compute_induction_score and compute_previous_token_score were implemented from scratch.

Formation of Previous Token Heads

Previous Token Heads concentrate in layers 0–4. Head L0H3 shows a steady plateau at 0.20 from 14,000 steps onward. Dynamics stabilize by the end of training (20,000 steps, ~5.5 epochs).

Google AdInline article slot

Distribution across layers and heads:

  • L0H3: 0.20 (maximum)
  • Layers 0–4: dominant values
  • Later layers: minimal contribution

This points to a fundamental mechanism essential for next-token prediction, regardless of data complexity.

Analysis of Induction Heads

Induction Heads don't fully develop: peak of 0.0537 at L6H4 around 15–16,000 steps, then a drop-off. No head exceeds 0.05, aligning with the low repetition statistics in TinyStories.

Google AdInline article slot

Key observations:

  • Peak in a mid-layer (L6).
  • No stable plateau—indicating instability.
  • Comparison to Previous Token: 4x lower score.

Hypothesis: The dataset's simplicity makes Induction mechanisms redundant for minimizing loss.

Sparse Autoencoder Results

SAE trained on activations from layer 6 (L6H4). Active features per token: 20–35. Dead features: ~3% (228 out of 8192) thanks to L1 regularization.

Google AdInline article slot

Two dominant features:

  • F#8107: syntactically predictable context, activating on first tokens (difference from repeats: 0.04–0.12).
  • F#635: general contextual pattern at predictable positions.

No pure Induction features detected—confirming the weak score.

Dataset Impact on Mechanisms

TinyStories doesn't encourage In-Context Learning: the model relies on Previous Token Heads for low loss. On complex datasets (e.g., OpenWebText), Induction mechanisms should grow due to frequent repeats.

Next test plans:

  • Models of varying sizes on TPU.
  • TinyStories vs. OpenWebText comparison.
  • Measuring dependence on data statistics.

Key Takeaways

  • Previous Token Heads form universally (score ~0.20 in early layers).
  • Induction Heads are weak on simple data (max 0.05, unstable).
  • SAE is effective: 20–35 active features, 3% dead features.
  • No pure Induction features—F#8107/635 tied to syntax/context.
  • Hypothesis: Dataset complexity drives ICL development.

— Editorial Team

Advertisement 728x90

Read Next