---
license: apache-2.0
language: [en]
tags: [nla, natural-language-autoencoders, qwen2.5, activation-vectors, warm-start]
---

# NLA Qwen2.5-7B L20 warm-start data — Sonnet-4.6 "matryoshka" explanations + activations

Re-warm-start dataset for the [Natural Language Autoencoders](https://github.com/kitft/natural_language_autoencoders)
Qwen2.5-7B (layer-20) AV/AR pair. Pairs **Qwen2.5-7B-Instruct layer-20 residual-stream
activations** with the **Claude Sonnet-4.6 explanations** from
[`ceselder/nla-matryoshka-warmstart-sonnet46`](https://huggingface.co/datasets/ceselder/nla-matryoshka-warmstart-sonnet46).

The source dataset is **text-only** (explanations keyed by `custom_id`, no vectors). This
dataset adds the missing activations: for each `av-*`/`ar-*` row the activation is
Qwen2.5-7B-Instruct's layer-20 hidden state at the **last token of `input_text`**.
`input_text` is exactly `token_ids[:position]` — verified: it retokenizes to the `custom_id`
position count for 100% of rows — so this is reproducible from the text alone, **no source
corpus needed** (the original Ultra-FineWeb slice is no longer hosted / has drifted).

## Files
| file | rows | use |
|---|---|---|
| `av_sft.parquet` | 219992 | actor (AV) warm-start SFT — train |
| `av_eval.parquet` | 5007 | actor eval holdout (document-level) |
| `ar_sft.parquet` | 219600 | critic (AR) warm-start SL — train |
| `ar_eval.parquet` | 5008 | critic eval holdout (document-level) |
| `base_av.parquet` | 224999 | master: vectors+explanation+text+custom_id (all av) |
| `base_ar.parquet` | 224608 | master: vectors+explanation+text+custom_id (all ar) |

Each parquet ships a `.nla_meta.yaml` sidecar. SFT/eval parquets are stage-3 format
(drop-in for `configs/actor_sft.sh` / `critic_sft.sh`). The `base_*` masters carry the
**raw activation vectors** (`norm="none"`) + `api_explanation` + `detokenized_text_truncated`
+ `custom_id`; run `nla.datagen.stage3_build` on them to regenerate SFT parquets with other
templates, or to re-split.

## Holdout
~10k samples held out for eval (≈5k av + ≈5k ar), split at the **document level**
(whole docs, seed 42) so no training position shares a document with an eval position.

## Provenance
- base model `Qwen/Qwen2.5-7B-Instruct`, layer 20, d_model 3584, raw vectors (`norm="none"`)
- explanations `claude-sonnet-4-6` (av-*/ar-* succeeded rows of the matryoshka dataset)