AmiriHayes/sae_token_programs

Programs for GPT-2 SAE features

A readable Python program for each of 301,090 SAE features of GPT-2 small — all 13 gpt2-small-res-jb checkpoints (blocks.0-11.hook_resid_pre plus blocks.11.hook_resid_post), 24,576 features each.

Each program is an ordinary function that takes the tokens seen so far and returns whether that feature fires:

def prog_L06F08045(input_tokens):
    """3 token families, 4 surface forms (train F1 of the uncapped optimum: 0.958)."""
    fire = {
        ' George', 'George',   # george  (3,592 firings)
        ' Osama',              # osama   (214 firings)
        'ORGE',                # orge    (28 firings)
    }
    return ('george-family', 1) if input_tokens[-1] in fire else ('no match', 0)

Fitted on 289,658 OpenWebText sequences (36,786,566 read positions), scored on a structurally disjoint 10,000-sequence test split. No LLM was used to write these — the token sets are chosen by greedy marginal F1.

Layout

path	what
programs/layer{00..12}/	the programs, 100 per file, in thousand-blocks
scores/layer{00..12}.csv	per-feature test F1, precision, recall, token count
synthesis_attempt_1/	387 LLM-written layer-6 programs — an earlier attempt, kept for comparison
activation_value_program/	one program that predicts the activation value, not just whether it fires

Results

Median test F1 by layer, and the single best program in each:

	L0	L1	L2	L3	L4	L5	L6	L7	L8	L9	L10	L11	L12
median	.98	.86	.80	.73	.65	.55	.41	.30	.23	.19	.17	.14	.13
best	1.00	1.00	1.00	1.00	.98	.99	.97	.96	.95	.95	.94	.93	.90

The median falls steeply with depth while the best barely moves. Depth doesn't cap what a token-only rule can reach — it changes how many features are reachable by one.

Caveats

• scores/ reports test F1; the docstring inside each program reports the train optimum. They are different numbers.
• Layers 0–3 have fewer than 24,576 programs because some features fire only at position 0 (attention-sink features) and are excluded from fitting.