Circuit atlas networks

The networks behind the entries of the Mechanistic Interpretability Circuit Atlas (https://github.com/jc1122/mechanistic-interpretability-circuit-atlas): a frozen hand-built transformer ("m2"), a frozen hand-written menu of encodings, and a trained one-hidden-layer MLP on top (rung PART). Code to rebuild m2 and the menu, train and analyse: the atlas repository's code/ (MIT). Weights and cards: CC BY 4.0.

Folder Entry Networks
_shared/m2/ hand-built base (rung HAND), GPT-NeoX parameter names, config.json 1
modular-addition-mlp/ (x + y) mod 59 10 seeds, train fraction .4
count-mod-5/ count of a letter mod 5 10 seeds x fractions .4/.6/.8
compare-counts/ which letter is more frequent 10 seeds x fractions .4/.6/.8
open-bracket-count/ number of ( in a bracket line 10 seeds x fractions .4/.6/.8
base-conversion/ hex/binary/decimal conversion of 0..255 (rungs HAND, PART, TRAIN; own layout, see its card) 83
decimal-hex-addition/ a + b, 3 digits, base 10 and 16, formats A/B/C (rungs HAND, PART, TRAIN; own layout, see its card) 42
binary-addition/ a + b, 6 bits, formats A/B/C (rungs HAND, PART, TRAIN; own layout, see its card) 62
subtraction/ a - b, base 10 (3 digits) and base 2 (6 bits), formats A/B/C (rungs HAND, PART, TRAIN; own layout, see its card) 112
binary-multiplication/ a * b, 6 bits, formats A/B (rungs HAND, PART, TRAIN; own layout, see its card) 41
long-division-square-root/ binary long division and long-hand square root, 8-bit and wide ranges, formats A/B (rungs HAND, PART, TRAIN; own layout, see its card) 202
dfa-group-automata/ parity, Z3 and S3 group automata (state tracking), formats A/B (rungs HAND, PART, TRAIN; own layout, see its card) 223
gf2-matrix-product/ C = A B over GF(2), 3x3 bit matrices, formats T (trace)/A (rungs HAND, PART, TRAIN; own layout, see its card) 15
gf2-elimination/ A x = b over GF(2) by Gauss-Jordan, m = 3 and 4, formats T (scratchpad)/A (rungs HAND, PART, TRAIN; own layout, see its card) 35
router-addition/ a + b, base 2 and 10, formats AM/AL/I/R (router), length generalisation (rungs HAND, PART, TRAIN; own layout, see its card) 105
multiplication-base-order/ x * y, x, y < 256, answer only and running-sum scratchpad, bases 2/10/16, MSB/LSB-first, curriculum mixes (rung TRAIN; own layout, see its card) 43
gcd-euclid/ gcd, modular inverse, fraction reduction; formats A (answer only)/T (subtractive-Euclid trace) (rungs HAND, PART, TRAIN; own layout, see its card) 39
modular-arithmetic-mod-97/ a * b and a + b mod 97, residue tokens and binary (rungs HAND, PART, TRAIN; own layout, see its card) 35
modular-arithmetic-formats/ a * b and a + b mod 97, residue tokens, final run (rungs HAND, TRAIN; own layout, see its card) 20
chess-krk/ king + rook vs king mate, fixed 8-rule fence policy, formats A/G/C (rungs PART, TRAIN; own layout, see its card) 41
chain-rule-differentiation/ symbolic derivative of prefix expressions (powers, sin, cos, exp, +, *), formats A (answer only)/T (register rewrite) (rungs HAND, PART, TRAIN; own layout, see its config) 30
polynomial-differentiation/ integer polynomial differentiation (power rule), formats A (answer only)/T (scratchpad) (rungs HAND, PART, TRAIN; own layout, see its card) 60
polynomial-add-mul/ integer polynomial addition and multiplication, degree 3, formats A (answer only)/T (scratchpad) (rungs HAND, PART, TRAIN; own layout, see its card) 59
output-order-base-conversion/ 16-bit decimal/binary conversion, answer MSB- or LSB-first, answer only (A) and scratchpad (B) (rungs HAND, PART, TRAIN; own layout, see its card) 95
symbolic-algebra-atoms/ symbol-level polynomial collection, binomial expansion and linear equations on (side, power, coefficient) atoms, formats A/T (rungs HAND, PART, TRAIN; own layout, see its card) 72
polynomial-integration/ integer polynomial integration with reduced-fraction coefficients, formats A (answer only)/T (subtractive-Euclid scratchpad) (rungs HAND, PART, TRAIN; own layout, see its card) 40
grid-bfs/ shortest-path distance on 6x6 mazes, formats A (answer only)/T (reached-set map scratchpad) (rungs HAND, PART, TRAIN; own layout, see its card) 18
compose-arithmetic-registers/ base-2 register scratchpad: STEP (ADD/SHL/REV on tagged registers) and its compositions SUM/MUL/POW, formats O/S (rungs HAND, PART, TRAIN; own layout, see its card) 14
dpll-search/ DPLL search on 3-SAT (n 4-6, m 10-25), formats A (answer only)/T (solver-state scratchpad) (rungs HAND, PART, TRAIN; own layout, see its card) 44
chess-mate-in-one/ mate-in-one candidate selection (bare black king, K 1-16 candidates), formats A (answer only)/T (escape-count scratchpad) (rungs HAND, PART; own layout, see its card) 24
linear-systems-finite-fields/ A x = b over F5/F7 by Gauss-Jordan, m = 2/3, formats A (answer)/T (matrix updates)/S (microstep scratchpad) (rungs PART, TRAIN; own layout, see its card) 158
quadratic-roots/ integer roots/discriminant of x^2 + bx + c: parent formats A/S/T and an addressed-NAND register executor S_exec (rungs HAND, PART, TRAIN; own layout, see its card) 33
crc-remainder/ GF(2) polynomial remainder N mod D (degree-4 divisors), formats A (answer only)/T (quotient + remainder scratchpad) (rungs HAND, PART, TRAIN; own layout, see its card) 38

Per entry: s<seed>_f<frac*100>.safetensors (W1, b1, W2, b2 float32; test_idx int64 = held-out cells), inputs.safetensors (X = frozen menu rows at the answer position, one per cell; uv = the two integers; lab = labels), config.json (shapes, menu block slices, optimiser) and manifest.json (seed, rung, fraction, circuit type, accuracies, sha256). Forward pass: logits = W2 relu(X W1^T + b1) + b2; accuracy on test_idx reproduces acc_test in the manifest exactly. No pickles. Each entry folder has its own card (README.md).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support