sandrajyluo/nlp-final-project-activations

Paired teacher/student activations on SVAMP and LOGIC-701

Activations and per-question results from running inference/run_eval_with_hooks.py on the nlp-final-project repo.

• Teacher: unsloth/Llama-3.2-1B-Instruct (mirror of meta-llama/Llama-3.2-1B-Instruct), explicit chain-of-thought via chat-template prompting, greedy decoding.
• Student: zen-E/CODI-llama3.2-1b-Instruct (LoRA + projection on the same base), 6 latent thought iterations, greedy decoding.
• Hook position (teacher): residual stream at the last prompt token, every layer.
• Hook position (student): residual stream at every latent step, every layer.

Files

Path	Shape	Dtype	Notes
svamp_teacher/activations.pt	(1000, 17, 2048)	bf16	(N, num_layers + 1 emb, hidden)
svamp_student/activations.pt	(1000, 6, 17, 2048)	bf16	(N, latent_steps, num_layers + 1 emb, hidden)
logic701_teacher/activations.pt	(701, 17, 2048)	bf16
logic701_student/activations.pt	(701, 6, 17, 2048)	bf16
*/results.json	—	text	per-question {idx, question, gold, pred, correct, response}

num_layers + 1 = 17 includes the embedding output (index 0) plus 16 decoder block outputs (1..16).

Confusion (teacher × student)

SVAMP (N=1000) — teacher 0.622, student 0.608, agreement 0.708

	Student correct	Student incorrect
Teacher correct	469	153
Teacher incorrect	139	239

LOGIC-701 (N=701) — teacher 0.233, student 0.031, agreement 0.745

	Student correct	Student incorrect
Teacher correct	3	160
Teacher incorrect	19	519

Reproduction

git clone git@github.com:sandraluo22/nlp-final-project.git
cd nlp-final-project/inference
# H100 / single GPU; takes ~10 min total
bash run_sweep.sh

LOGIC-701 student accuracy is artificially low because ~96% of free-text responses contain no parseable option number. A log-likelihood scorer over the five options would give a more honest comparison.