README.md

---
license: mit
task_categories:
  - feature-extraction
tags:
  - mathlib
  - lean4
  - theorem-embeddings
---

# vector_behavior — data bundle

Companion data for https://github.com/david-busbib/vector_behavior-

Four theorem embedding spaces over the same 99,552 Mathlib theorems, restricted to
the rows the study uses and written **in pool order**, float32. Also carries the
proof-dependency graph, reduced to what the scripts actually read.

| directory | space | dim | size |
|---|---|---|---|
| `embeddings_mistral/` | `e5_mistral` — intfloat/e5-mistral-7b-instruct | 4096 | 1.6 GB |
| `embeddings_corpus/` | `e5_large` — intfloat/e5-large-v2 | 1024 | 407 MB |
| `embeddings_fixed/` | `e5_large_stmt` — same model, different text source | 1024 | 407 MB |
| `embeddings_deepseek/` | `deepseek_prover` — **degenerate, contrast only** | 4096 | 1.6 GB |
| `results_shared/` | pool.json, subset.json, neighbors.npz | | 27 MB |
| `graph_uses.json.gz` | proof dependency graph, 55,575 theorems | | 2.4 MB |

## Setup

```bash
git clone https://github.com/david-busbib/vector_behavior-.git
cd vector_behavior-
pip install -r requirements.txt

huggingface-cli download davidbusbib/vector-behavior-data \
    --repo-type dataset --local-dir vb_data

cp vb_data/results_shared/* results/shared/
export VB_DATA=$PWD/vb_data

python scripts/19_read_mutual.py
```

### Pulling only what you need

Each space is ~0.4-1.6 GB. To take one:

```bash
huggingface-cli download davidbusbib/vector-behavior-data \
    --repo-type dataset --local-dir vb_data \
    --include "embeddings_corpus/*" "results_shared/*" "graph_uses.json.gz" "to_additive.json"
```

Scripts filter themselves to whichever spaces are present, so a partial download
runs the whole pipeline over the subset you took.

## Notes

* Rows are pool-ordered, so `pool["rows"][space]` is the identity. This is why the
  arrays are smaller than the originals (99,552 rather than 180,907 / 104,050 rows)
  and why no script needs changing.
* float32 throughout. float16 moves the numbers.
* `deepseek_prover` is degenerate — 81% of its variance sits in one direction after
  centring, and it was extracted under fp16. It is in the study as a contrast, never
  as evidence. See `RESULTS.md`.
* Proofs and the dependency graph are **held back** as validation in this project.
  `graph_uses.json.gz` is here so Phase 3 can be rerun; it is not an encoder input.