dataset_id: av_sft_Qwen3-8B_L24_6a22e324__av_sft_raw__explained__shuf42
stage: av_sft
row_count: 247261
extraction:
  base_model: Qwen/Qwen3-8B
  d_model: 4096
  layer_index: 24
  norm: none
  corpus: /workspace-vast/celeste/nla-data/finefineweb_100k.parquet
  corpus_slice:
    start: 0
    length: 100000
  positions_per_doc: 10
kind: nla_dataset
schema_version: 1
keep_debug_metadata: true
tokens:
  injection_char: "\u320E"
  injection_token_id: 149705
  injection_left_neighbor_id: 29
  injection_right_neighbor_id: 522
  critic_suffix_ids: null
prompt_templates:
  actor: 'You are a meticulous AI researcher conducting an important investigation
    into activation vectors from a language model. Your overall task is to describe
    the semantic content of that activation vector.


    We will pass the vector enclosed in <concept> tags into your context. You must
    then produce an explanation for the vector, enclosed within <explanation> tags.
    The explanation consists of 2-3 text snippets describing that vector.


    Here is the vector:


    <concept>{injection_char}</concept>


    Please provide an explanation.'
  critic: 'Summary of the following text: <text>{explanation}</text> <summary>'
api_summaries:
  model: claude-sonnet-4-6
  max_tokens: 300
  temperature: 1.0
  instruction_prompt: "A language model needs to predict what text comes next after\
    \ a snippet which will be presented to you shortly. Identify the 2-3 most important\
    \ features it would use for this prediction.\nFocus on what the language model\
    \ must be \"thinking about\" at the point where the provided text ends. You should\
    \ not need to 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.\nOrder features by what is most important for predicting the\
    \ next tokens. Each feature should consist of a concise ~10-20 word description.\
    \ Feel free to include specific textual examples inline.\n\nFeature types to consider\
    \ (as inspiration, not a rigid checklist):\n- Syntactic/structural constraints:\
    \ \"unclosed parenthesis requires matching close\"\n- Immediate semantic expectations:\
    \ \"list promised three items but only two given\"\n- Stylistic/register patterns:\
    \ \"formal academic tone maintained throughout\"\n- Narrative/argumentative momentum:\
    \ \"thesis stated, supporting evidence now expected\"\n- Domain/genre signals:\
    \ \"medical case history following SOAP format\"\n- Repetition/continuation patterns:\
    \ \"same phrase structure repeating with variations\"\n\nThe final feature must\
    \ describe the very end of the presented sequence: its role, what it's part of,\
    \ and immediate constraints on what follows.\n\nFormat \u2014 IMPORTANT: keep\
    \ to ~80-100 words total and ALWAYS close the tag:\n<analysis>\n[first feature\
    \ \u2014 include specific examples when relevant]\n[second feature]\n[final feature:\
    \ the last token, its role, immediate constraints]\n</analysis>\n\nText to analyze:\n\
    \n<begin_text>{text}<end_text>"
parent_datasets:
- av_sft_Qwen3-8B_L24_6a22e324__av_sft_raw__explained
created_at: '2026-05-16T19:56:05.443944+00:00'
created_by: nla.datagen.stage_shuffle
git_commit: 047eb8e
multi_input_slots: 16