File size: 3,553 Bytes
4c27541
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
pretty_name: WeightCode records and weights
tags:
  - self-improvement
  - llm-agents
  - continual-learning
  - model-editing
---

# WeightCode: records and weights

Experiment records and trained weights that accompany the WeightCode code release
([github.com/Gen-Verse/WeightCode](https://github.com/Gen-Verse/WeightCode)). The code repository is small and holds no
data. This dataset holds what a reader needs to audit the reported numbers and to rerun the learned programs.

## Layout

| Path | What it holds |
|---|---|
| `records/sessions/<id>/` | Text records of 88 LLM sessions: task, prompts, tool calls, transcript, attempts, patches, `final/` program and selection, audit and result files |
| `records/sealed/` | Numbers of record from the sealed reads: `registry.jsonl`, `manifests/`, `results/` (per-entry metrics and hashes, no test pixels) |
| `records/tables/` | Reference-stream, sealed-read, T07 and mechanism tables copied from the run store |
| `records/finals/` | The frozen finals lists, `finals_v1.json` to `finals_v6_secondary.json` (also in the GitHub repo) |
| `weights/instances/<id>/` | The 23 networks and instance specifications that the finals lists refer to (`model.npz`, `evidence.npz`, `model_spec.json`, `forward.py`, `task.json`) |
| `weights/session_finals/<id>/` | Final parameter deltas and edited models of the 21 sessions that the finals lists refer to |
| `MANIFEST.json` | Session origins, original directory of every instance, list of exclusions |
| `SHA256SUMS` | Checksum of every file except this card |
| `REDACTIONS.json` | The files in which two account identifiers were replaced, and what was replaced |

Session ids and instance ids are the ones used in the finals lists. `child_sha256` in a finals entry identifies the
program that was evaluated.

## Using it

This is not a drop-in experiment store. The code expects `$WEIGHTCODE_RUNS/sessions/<id>` and
`$WEIGHTCODE_PRIVATE/instances/<id>`. To rebuild that layout, link `records/sessions` and `weights/instances` to those
paths. Instances that were originally kept in other private subdirectories (for example `scan_instances` and
`e14_instances`) are all under `weights/instances` here, and `MANIFEST.json` gives their original location.

```python
from huggingface_hub import snapshot_download
snapshot_download("Boom5426/WeightCode", repo_type="dataset", local_dir="weightcode_data")
```

Verify a download with `sha256sum -c SHA256SUMS` from the dataset root.

## Redaction

On 2026-11-10 two identifiers of the authors' compute accounts were replaced in the text records: a personal login
name (now `<redacted-account>`) and a cluster project number (now `<redacted-project>`). `REDACTIONS.json` lists the
files and the number of replacements in each. Nothing else in those files changed: every number, hash, program and
weight is as recorded. A record that quotes the sha256 of one of these files quotes it as it was before the redaction,
so that quote no longer matches. `SHA256SUMS` gives the hashes of the files as published here.

## Not included

- Fixed-reference and fine-tuning trajectories (`.npy`, about 48 GB). They are regenerated by the scripts in the code
  repository.
- Qwen3 base weights, which are public from their original model pages.
- Public datasets (CIFAR-10, CIFAR-10-C, ImageNet and others) and sealed test pixels. The code repository contains the
  build scripts.
- Session working directories (`private/`, `scratch/`, `cwd/`) and arrays outside `final/`.

## License

MIT, the same as the code.