Learn by building: Chain-of-Thought (CoT) probing, in-context learning (ICL) emergence, induction-head scans, and distilled vs. base model comparisons.

Works out-of-the-box with sshleifer/tiny-gpt2 for CPU demos; swap to larger causal LMs for deeper dives.

pip install -e .
python -m inductra.runner icl --model sshleifer/tiny-gpt2 --ctx 8 16 32 64
python -m inductra.runner cot --model sshleifer/tiny-gpt2 --layer -1
python -m inductra.runner compare --base gpt2 --distilled distilgpt2

• ICL emergence grid — synthetic tasks (copy, reverse, modular add, parity) across context lengths; plot accuracy vs context.
• Induction heads — shifted-copy attention score for detecting induction-like heads.
• CoT probing — linear probe on residual streams to classify presence of CoT prefix (“Let’s think step by step”).
• Distilled vs base — quick accuracy delta on modular arithmetic.
• Hooks & probes — ~100 LOC each, minimal and hackable.

This repo pairs well with How to Become a Mechanistic Interpretability Researcher by Neel Nanda (Alignment Forum, 2025).

Distilled raw notes from the guide:

• Stage 1 (≤1 month): breadth-first basics, code a transformer, replicate toy mech interp tasks. Don’t wait—experiment early.
• Stage 2 (1–5 day projects): maximize info gain per unit time, run fast probes, debug aggressively.
• Stage 3 (1–2 week sprints): develop skepticism, set real baselines, and always write up. Public write-ups are the best credential.

Skill growth rates:

• Fast → coding, debugging, plotting.
• Medium → taste & prioritization, built by doing multiple projects.
• Slow → idea generation, emerges over months of exploration.

Practical tips: use LLMs as collaborators (they’re decent at mech interp tasks), don’t block on mentorship, and focus on “doing” over endless reading.

• Hook adapters for Llama/Mistral/Qwen
• Token-shifted causal tracing
• Per-head ranking plots
• GSM8K CoT token supervision
• Path patching experiments

MIT