File size: 1,153 Bytes
b9c5bae | 1 2 3 4 5 6 7 8 9 10 11 12 | ---
license: cc-by-4.0
tags: [mechanistic-interpretability, circuits, arithmetic, matrix-vector-product, scratchpad]
---
# integer-matvec
Atlas entry: entries/integer-matvec.md in https://github.com/jc1122/mechanistic-interpretability-circuit-atlas (mechanism, setup, limits). Task: y = Mv for a 3x3 integer matrix and vector with entries 0..7, each y_i as nine LSB-first bits. Format A emits the answers only; format T writes the row accumulator after every shifted partial product (scratchpad). Score: teacher-forced whole-output exact on each seed's disjoint 2,000-input test set.
Model: GPT-NeoX, 5 layers, 4 heads x 32, d_model 128, ReLU MLP 640, no LayerNorm, no positional encoding, vocab 15 (`config.json` -> `hf_config`).
Files: `hand_{A,T}` (hand-built HAND-a for A, HAND-b for T); `train_<fmt>_<round>_{best,final}_s<seed>` (rung TRAIN, full nets): T screening r1 s0-2, A retry r3 s0-2, T final run s0-9. PART research nets are not included. Per net: format, round, snapshot, seed, step, circuit type, `recorded_test`, `acc_test_fp32` (fp32 re-score), sha256: `manifest.json`.
34 networks (2 HAND, 32 TRAIN). No pickles. License CC BY 4.0.
|