File size: 9,826 Bytes
bab5617 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 | ---
license: apache-2.0
pretty_name: CSP-Atlas
size_categories:
- 100K<n<1M
tags:
- mechanistic-interpretability
- code
- python
- sparse-transformer
- circuits
- concept-extraction
language:
- code
---
# CSP-Atlas: Concept-Specific Neural Circuits in a Sparse Python Transformer
Frozen experimental artifacts for the paper *CSP-Atlas: Concept-Specific Neural Circuits in a Sparse Python Transformer* (Wilam, 2026). This repository holds the prompts, masks, and decomposition tables that the paper's figures and numbers are computed from; the model itself is `openai/circuit-sparsity` on the Hub.
- **Paper code:** <https://github.com/piotrwilam/AtlasCSP>
- **Model used:** [`openai/circuit-sparsity`](https://huggingface.co/openai/circuit-sparsity) — 8-layer sparse Python transformer, 2,048-dim MLP output per layer.
- **Concept space:** 106 Python concepts = 43 AST node types + 63 builtin objects.
- **Prompts:** 63,800 object prompts (1,276 pairs × 50 variations) + 2,848 checker + 2,861 token-checker.
- **Parameter sweep:** ε ∈ {0.001, 0.1, 0.5} × C ∈ {20%, 50%, 80%} = 9 settings.
All claims in the paper regenerate from these files via the released code (`pytest tests/test_paper_numbers.py`).
## Layout
```
CSP-Atlas/
├── 11A_object_prompts.parquet # 63,800 object prompts (the (AST × builtin) × 50 grid)
├── 11B_checker_prompts.parquet # 2,848 checker prompts (keyword present, concept absent)
├── token_checker_prompts.parquet # 2,861 token-checker prompts (broader set, all keyword contexts)
│
├── 12_object_activations.h5 # Raw MLP-output activations, object prompts (optional — derivable)
├── 12_checker_activations.h5 # Raw MLP-output activations, checker prompts (optional)
│
├── 13_object_masks_eps{ε}_cons{C}.h5 # 9 files — binarised + marginalised universal masks per (ε, C)
├── 13_checker_masks_eps{ε}_cons{C}.h5 # 9 files — binarised checker masks per (ε, C)
├── token_checker_masks.h5 # Token-checker masks (single setting: ε=0.5, C=0.8)
│
├── 14_summary.csv # 9 rows — global concept/token fractions per (ε, C)
├── 14_target_checker_eps{ε}_cons{C}.csv # 9 files — per-object × per-layer "co / sh / to" decomposition
│
├── relaxed_modularity_scores.csv # §6.1 — significant-layer count per concept × trim level p ∈ {0, 0.05, 0.1, 0.25}
├── relaxed_modularity_detail.csv # §6.1 — full per-layer Jaccard trim-mean values
│
├── token_independence_summary.csv # Per-object: concept_fraction, token_fraction, A_size, A∩B, A\B, B\A
├── token_independence_detail.csv # Per-object × per-layer breakdown
└── token_independence_no_keyword.csv # The 48 tokenless concepts (no checker confound applies)
```
## File schemas
### HDF5 mask files (`13_*.h5`, `token_checker_masks.h5`, `universal_106x50.h5`)
Layer-major group layout. Datasets are boolean vectors of length 2,048 (one bit per MLP-output dimension):
```
/metadata
.attrs: {epsilon, consistency_thresh, n_layers=8, n_pairs=1276}
/ast_nodes — array of 43 AST node names
/builtin_objs — array of 63 builtin object names
/universal_masks
/layer_0/{concept_name} — (2048,) bool
/layer_1/{concept_name} — (2048,) bool
…
/layer_7/{concept_name} — (2048,) bool
/pair_masks
/layer_0/{ast}__{builtin} — (2048,) bool (e.g. AnnAssign__abs)
…
/metrics
aggregate statistics per layer (sizes, intersection counts)
```
Concept names use the convention `ast__<NodeType>` or `builtin__<object>` (double underscore).
### Prompt parquets
`11A_object_prompts.parquet` (63,800 rows):
| column | type | meaning |
|---|---|---|
| `ast_node` | str | AST node type (e.g. `For`) |
| `builtin_obj` | str | Builtin name (e.g. `bytearray`) |
| `variation_id` | int | 0..49 (50 variations per pair) |
| `prompt_text` | str | The Python snippet |
| `sequence_loss` | float | Model's per-token loss on the snippet (used for filtering) |
| `token_length` | int | Tokenizer output length |
| `ast_verified` | bool | True if `ast.parse()` round-trips with the target node present |
`11B_checker_prompts.parquet` (2,848 rows) and `token_checker_prompts.parquet` (2,861 rows):
| column | type | meaning |
|---|---|---|
| `object` | str | Target concept (e.g. `ast__Break`) |
| `keyword` | str | The bare token under test (e.g. `break`) |
| `variation_id` | int | 0..49 |
| `prompt_text` | str | Python that contains the keyword but excludes the concept (string literal, comment, var name, …) |
### Decomposition CSVs (`14_target_checker_*.csv`)
One row per testable concept (58 of them), 8 layer columns (`L0`–`L7`). Each cell encodes the three-way decomposition as a string:
```
co / sh / to
```
— concept-only `|A \ B|`, shared `|A ∩ B|`, token-only `|B \ A|`. From these the concept fraction `co / (co + sh)` is recovered.
### Global summary (`14_summary.csv`)
One row per (ε, C) setting. Columns: `epsilon, consistency, n_testable, n_ast, n_builtin, mean_concept_fraction, mean_token_fraction, ast_concept_fraction, ast_token_fraction, blt_concept_fraction, blt_token_fraction, mean_A_size, mean_B_size`.
### Modularity (`relaxed_modularity_scores.csv`)
| column | meaning |
|---|---|
| (index) | concept name |
| `type` | `ast` or `builtin` |
| `p=0.0`, `p=0.05`, `p=0.1`, `p=0.25` | Number of layers (0–8) at which the concept's mean Jaccard to its k-nearest-neighbour shell, trimmed at trim level p, exceeds the permutation null at significance 0.05. The paper's §6.1 Break-at-top result is `p=0.0 == 3`. |
`relaxed_modularity_detail.csv` is the un-aggregated per-layer per-trim-level Jaccard means feeding the score table.
### Token-independence (`token_independence_*.csv`)
`summary.csv` is one row per testable object: `object, type, keyword, concept_fraction, token_fraction, A_size, A_and_B, A_only, B_only` (averaged across layers).
`detail.csv` breaks the same statistics out by `(object, layer)`.
`no_keyword.csv` is the 48 concepts (19 tokenless AST + 29 tokenless builtins) for which no checker confound applies — included so downstream users can see the full 106-concept universe.
## Quickstart
The dataset is normally accessed through the [AtlasCSP code repository](https://github.com/piotrwilam/AtlasCSP), which contains loaders that auto-download missing files from this Hub mirror:
```python
from csp_atlas.io import (
load_universal_masks, # 13_object_masks_*.h5
load_pair_masks, # 13_object_masks_*.h5
load_concept_inventory, # the 43 + 63 names
load_summary, # 14_summary.csv
load_target_checker, # 14_target_checker_*.csv
load_modularity_scores, # relaxed_modularity_scores.csv
load_prompts, # 11A / 11B / token_checker
)
# Universal mask set at the paper's reference setting:
masks = load_universal_masks(eps=0.5, cons=0.8) # dict[concept] -> dict[layer] -> set[int]
ast, builtin = load_concept_inventory(eps=0.5, cons=0.8) # 43, 63
prompts = load_prompts(kind="object") # the 63,800-row DataFrame
```
Direct download with `huggingface_hub`:
```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="piotrwilam/CSP-Atlas",
filename="13_object_masks_eps0.5_cons0.8.h5",
repo_type="dataset",
)
```
Or with `datasets` (parquets only):
```python
from datasets import load_dataset
ds = load_dataset("piotrwilam/CSP-Atlas", data_files="11A_object_prompts.parquet")
```
## Reproducing the paper
The paper's claims are locked in `tests/test_paper_numbers.py` in the code repo. With the dataset in place at `$CSP_ATLAS_DATA_ROOT` (or the default local mirror) and a uv venv:
```bash
git clone https://github.com/piotrwilam/AtlasCSP
cd AtlasCSP
uv sync
export CSP_ATLAS_DATA_ROOT=/path/to/CSP-Atlas
pytest tests/test_paper_numbers.py -v
python experiments/fig1_atomicity_dendrogram.py
```
The single dendrogram figure (Figure 1) is produced by `experiments/fig1_atomicity_dendrogram.py`, driven by `configs/paper/figure1_atomicity_dendrogram.yaml`.
## Generation pipeline
The four-stage extraction (`11_` → `12_` → `13_` → `14_`) is documented in [`circuits/README.md`](https://github.com/piotrwilam/AtlasCSP/blob/main/circuits/README.md) in the code repo:
| Stage | Step | Output |
|---|---|---|
| 11 | Prompt generation (`circuits/prompts/`) | `11A`, `11B`, `token_checker` parquets |
| 12 | Forward-pass extraction (`circuits/extraction/`) | `12_*_activations.h5` (last-token MLP outputs, 8 layers) |
| 13 | Binarisation + marginalisation | `13_object_masks_*.h5`, `13_checker_masks_*.h5` per (ε, C) |
| 14 | Decomposition vs checker masks | `14_summary.csv`, `14_target_checker_*.csv` |
The `relaxed_modularity_*.csv` and `token_independence_*.csv` are downstream analyses computed on stage 13/14 outputs.
## What is *not* in this repo
- **Model weights** — `openai/circuit-sparsity` is loaded from its own Hub repo at runtime; not duplicated here.
- **Pre-paper development snapshots** — checkpoints/, earlier 115-concept experiments, exploratory runs.
- **Figure PDFs** — the figure files for the paper live in the AtlasCSP code repo under `results/`.
## Citation
```bibtex
@article{Wilam2026CSPAtlas,
title = {CSP-Atlas: Concept-Specific Neural Circuits in a Sparse Python Transformer},
author = {Wilam, Piotr},
year = {2026},
url = {https://github.com/piotrwilam/AtlasCSP}
}
```
## License
Apache-2.0 — see `LICENSE`. The model `openai/circuit-sparsity` is distributed under its own licence on its own Hub repo and is not redistributed here.
|