syvb/nla-warmstart-explanations-finefineweb-sonnet46

NLA Warmstart Explanations — FineFineWeb / Sonnet 4.6 (test run)

A small test/sample run (8,398 rows) of supervised-warmstart explanations for Natural Language Autoencoder (NLA) training, generated with the Claude Sonnet 4.6 Message Batches API.

Each row pairs a pretraining-text snippet with a natural-language description of the features a language model would use to predict the next tokens at the end of that snippet. In NLA training these descriptions are the supervised target for the activation verbalizer (AV) and the input text for the activation reconstructor (AR).

This is a dry-run before a full-scale generation — published to validate the pipeline (prompt → batch generation → extraction → release) end to end.

Columns

Column	Type	Description
detokenized_text_truncated	string	The source text snippet (prefix up to a sampled token position). Cache key.
api_explanation	string	Sonnet 4.6 explanation: 10 next-token-prediction features, \n\n-separated
doc_id	int	Source document id
n_raw_tokens	int	Token position in the source doc where the snippet was truncated
activation_layer	int	Layer the paired activation was taken from (24) in the source pipeline

How it was generated

• Model: claude-sonnet-4-6, Message Batches API (50% pricing), max_tokens=300, temperature=1.0
• Source texts: snippets from m-a-p/FineFineWeb, drawn from the Qwen3-8B / layer-24 NLA warmstart corpus slice (100k docs × 10 positions). The explanation depends only on the text, not the model — so it is reusable across base models.
• Yield: 8,399 generated, 8,398 kept (1 dropped for failing the <analysis> extraction / <2-feature filter).
• Cost: $21.60 (input 8.0M tok, output 1.3M tok, batch-discounted).

Intended use — reusable explanation cache

The explanations are keyed on detokenized_text_truncated, so they plug directly into the NLA datagen pipeline's --cache-from to skip re-calling the API for texts already explained:

python -m nla.datagen.stage2_api_explain \
    --input splits/av_sft_raw.parquet \
    --output splits/av_sft_explained.parquet \
    --cache-from nla-warmstart-explanations.parquet

Prompt

Generation prompt (click to expand)

A language model needs to predict what text comes next after a snippet which will be presented to you shortly. Identify the 10 most important features it would use for this prediction.
Focus on what the language model must be "thinking about" at the point where the provided text ends. You should not need reference the fact that the text is truncated/incomplete/a prefix: the language model is causal, so only sees the prefix to what it predicts and this is implicit.
Order features by their overall importance for predicting the next tokens.
Each feature should consist of a 5-10 word description and should make sense by itself. When describing patterns, feel free to:

Note when patterns repeat, mirror, or continue from earlier in the text
Show how the same pattern manifests in multiple forms
Use probabilistic language ("often", "typically", etc.) when the model faces uncertainty
Feel free to include specific - and relevant - textual examples inline.
Track both what tokens are expected AND what contextual patterns determine those expectations

Feature types to consider (as inspiration, not a rigid checklist):

Syntactic/structural constraints: "unclosed parenthesis requires matching close before sentence ends"
Immediate semantic expectations: "list promised three items, third now required"
Stylistic/register patterns: "formal academic tone, passive voice and latinate vocabulary"
Narrative/argumentative momentum: "thesis just stated, supporting evidence now expected"
Logical/causal dependencies: "causal premise established, economic consequence must follow"
Domain/genre signals: "medical case history in SOAP format, now Assessment section"
Discourse/dialogue context: "speaker interrupted mid-argument, same thought continues"
Repetition/continuation patterns: "same phrase structure repeating, like 'from a X perspective'"
Distributional expectations: "-ing verb typically followed by noun phrase here"
Epistemic/meta-textual stance: "hedging with 'may' and 'possibly' signals continued uncertainty"

Additionally:
Be specific and precise. Consider how the model tracks patterns across multiple levels simultaneously-surface forms, semantic content, genre conventions, and sequential dependencies. Features can describe both immediate next-token constraints and longer-range structural expectations.
Only write one newline between features.
Always write exactly 10 features.

Format (use up to 50-100 words total, MAXIMUM 10 words per item):
<analysis>
[first feature (most important, describes final token)]
[second feature]
[...]
[final feature (least important)]
</analysis>

Text to analyze:

<begin_text>{text}<end_text>

Limitations

NLA-style explanations are known to confabulate — they can contain specific claims about the text that are false, while remaining thematically plausible (see the NLA paper). Treat individual feature claims as noisy; signal is strongest in themes repeated across a context.

Provenance & license

Text snippets derive from FineFineWeb (m-a-p) and remain subject to its terms. Explanations are generated by Claude Sonnet 4.6. Released for research use.