| --- |
| license: other |
| license_name: ai2-impact-lmr |
| extra_gated_prompt: >- |
| This dataset is derived in part from AI2's WildChat corpus, which is released |
| under the AI2 ImpACT License and is itself gated. By requesting access you |
| confirm that you accept WildChat's terms and will not redistribute the |
| conversation text contained here. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| I accept the WildChat (AI2 ImpACT) terms: checkbox |
| I will not redistribute the conversation text: checkbox |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - interpretability |
| - natural-language-autoencoder |
| - activation-verbalization |
| --- |
| |
| # NLA warmstart data (Qwen3-8B, layer 24) |
|
|
| Supervised-finetuning data for the two active natural-language-autoencoder (NLA) |
| families. An NLA is an encoder/decoder pair over a language model's residual stream: |
| a **verbalizer (AV)** turns one activation vector into English, and a |
| **reconstructor (AR)** maps that English back to the activation. These are the |
| warmstart sets the AV and AR are trained on before RL. |
|
|
| Companion weights (public, ungated): |
| [`asher577/nla-warmstarts`](https://huggingface.co/asher577/nla-warmstarts) · |
| [`asher577/nla-rl-ckpts`](https://huggingface.co/asher577/nla-rl-ckpts) |
|
|
| ## Why this repo is gated |
|
|
| Roughly half the 8-bullet data and **all** of the focus data derive from AI2's |
| **WildChat**, which is gated and licensed under the AI2 ImpACT License. The |
| `detokenized_text_truncated` column contains that conversation text verbatim. Access |
| here is gated to mirror WildChat's own terms — please honour them. |
|
|
| The remaining ~51% of the 8-bullet data derives from **FineFineWeb** (public web text). |
|
|
| ## Contents |
|
|
| | path | rows | what | |
| |---|---|---| |
| | `free8/av_sft.parquet` | 131,312 | 8-bullet verbalizer SFT (~49% WildChat / 51% FineFineWeb) | |
| | `free8/ar_sft.parquet` | 66,047 | 8-bullet reconstructor SFT | |
| | `focus/av_sft.parquet` | 249,746 | focus 4-aspect verbalizer SFT (100% WildChat) | |
| | `focus/ar_sft.parquet` | — | focus reconstructor SFT | |
|
|
| Every parquet ships with its `<name>.nla_meta.yaml` **sidecar, which is required** — |
| it carries the injection token ids, prompt templates, `injection_scale` and |
| `mse_scale` that the loaders assert against the live tokenizer at startup. Keep the |
| sidecar next to the parquet. |
|
|
| ## Schema |
|
|
| | column | type | meaning | |
| |---|---|---| |
| | `prompt` | list of chat messages | the instruction given to the model (contains the injection marker; **not** the source text) | |
| | `response` | string | the gold explanation to be learned | |
| | `activation_vector` | fixed_size_list<float>[4096] | the Qwen3-8B layer-24 residual-stream activation, **raw / unnormalized** | |
| | `n_raw_tokens` | int | tokens of context behind the activation | |
| | `activation_layer` | int | 24 throughout | |
| | `doc_id` | string | `<corpus parquet>:<split>:<row>` — the source row | |
| | `detokenized_text_truncated` | string | the source text the activation was taken from | |
|
|
| Activations are stored **raw**: normalization happens at injection time |
| (`injection_scale`) and at loss time (`mse_scale`), both read from the sidecar. |
|
|
| ## The two families |
|
|
| - **free8 (8-bullet)** — a free-form compositional verbalizer: 8 bullets whose |
| per-bullet reconstructions sum to the activation. Half web text, half chat. |
| - **focus (4-aspect)** — a WildChat verbalizer with four designated components: |
| provenance/context, persona, user information, and model goal. |
|
|
| Both target Qwen3-8B at layer 24. |
|
|
