File size: 2,917 Bytes
0d71050
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: odc-by
configs:
- config_name: av_sft
  data_files:
  - split: train
    path: av_sft_train.parquet
  - split: validation
    path: av_sft_val.parquet
- config_name: ar_sft
  data_files:
  - split: train
    path: ar_sft_train.parquet
  - split: validation
    path: ar_sft_val.parquet
tags:
- interpretability
- activations
---

# EasyNLA warmstart data (non-compositional, Qwen3-8B layer 24)

Supervised warm-start data for [EasyNLA](https://github.com/asherps/EasyNLA) —
train a **natural-language autoencoder** on Qwen3-8B: an activation-verbalizer
(AV) that explains a layer-24 residual activation in natural language, and an
activation-reconstructor (AR) that maps the explanation back to the activation.

Each row carries one **raw** (unnormalized) layer-24 activation captured while
Qwen3-8B read a FineFineWeb document (position ≥ 50 tokens), plus a
Claude-written "gold" explanation of what the model was representing there.

| config | split | rows | docs |
|---|---|---|---|
| `av_sft` | train | 363,961 | 36,734 |
| `av_sft` | validation | 7,307 | 737 |
| `ar_sft` | train | 363,988 | 36,734 |
| `ar_sft` | validation | 7,405 | 747 |

## Columns

- `prompt` — chat messages for the AV (with the `<INJECT>` marker placeholder)
  in `av_sft`; the AR critic prompt text in `ar_sft`
- `response` (`av_sft` only) — the gold explanation the AV imitates
- `activation_vector` — raw layer-24 residual, `list<float32>[4096]`
- `doc_id`, `n_raw_tokens`, `activation_layer`,
  `detokenized_text_truncated` (the source text up to the extraction position)

Each parquet has a matching `<name>.parquet.nla_meta.yaml` **sidecar** — the
contract EasyNLA trainers assert at startup (marker token IDs, prompt
templates, `d_model`, scales). Keep them next to the parquets.

## Splits

Document-level and deterministic: a doc goes to **validation** iff
`zlib.crc32(doc_id) % 1000 < 20` (`nla/val_split.py::is_val_doc(doc_id, 20)`,
~2% of docs). No document's rows are split across train/validation, and the
`av_sft` / `ar_sft` configs come from disjoint document sets by construction
(stage-1 of the EasyNLA datagen pipeline).

## Use with EasyNLA

```bash
huggingface-cli download asher577/easynla-warmstart-data --repo-type dataset \
    --local-dir <data>

python -m nla.train_sft --mode av --base-ckpt Qwen/Qwen3-8B \
    --parquet <data>/av_sft_train.parquet --sidecar <data>/av_sft_train.parquet \
    --heldout-parquet <data>/av_sft_val.parquet --save-dir <ckpts>/av
python -m nla.train_sft --mode ar --base-ckpt Qwen/Qwen3-8B \
    --parquet <data>/ar_sft_train.parquet --sidecar <data>/ar_sft_train.parquet \
    --heldout-parquet <data>/ar_sft_val.parquet --save-dir <ckpts>/ar
```

Provenance: source text from FineFineWeb (ODC-By); explanations generated with
Claude (claude-sonnet-4-6); activations from Qwen/Qwen3-8B. The `doc_id`
strings are opaque keys from the generating pipeline.