DSv4-Flash REAP Wiki

Wiki pages

On this page

Resources

External review

From the DeepSeek-V4-Flash-0731 interpretability project

What text we fed the model

Every text we used, where it came from, and how much of it there is. All data is on HuggingFace.

Religious texts (the main corpus)

We gathered texts from 9 religious traditions plus a large collection of Christian literature. Here's the full inventory:

TextPassagesTokensDigit densityWhere it came from
KJV Bible1,1891,045,7760.0000%Project Gutenberg, verse numbers stripped
Christian literature (wave 1)1,26720,409,440~1.1%Project Gutenberg, 3,705 books across 32 topics
Christian literature (wave 2)2,295~37MvariesProject Gutenberg, extended topic set
Qur'an (Pickthall translation)115258,122variesProject Gutenberg
Book of Mormon30342,6140.000%Project Gutenberg
Bhagavad Gita1829,690lowProject Gutenberg
Tao Te Ching8113,852lowProject Gutenberg
Dhammapada2616,203lowProject Gutenberg
Analects (Confucius)2042,041lowProject Gutenberg
Upanishads321,669lowProject Gutenberg (too few for strong claims)
Total2,74922,179,407

The "digit density" column matters a lot, as it turned out. The Bible is the only text at exactly 0%, because we stripped verse numbers from it. That created the confound that led to our retracted "scripture detector" claim.

Experiment-specific text

For targeted experiments, we built custom corpora:

ExperimentPassagesTokensStatusWhat's in it
Exp 1: 12 Bible translations360613,344doneSame 30 passages in 12 English translations (KJV, WEB, ASV, YLT, BBE + NIV, ESV, NLT, NRSV, NASB, CSB, MSG). All with digits stripped.
Exp 4b: Commentary with quotes3515,156,659done351 chunks of Christian commentary that contain embedded Bible quotes. We annotated where the quotes are so we could compare routing on quotes vs. routing on commentary.
Exp 12: Digit pairs222437,900done111 pairs of the same text, one with digits and one with digits removed. 5 categories: dates, lists, statistics, scripture citations, misc numeric.
Exp 13: Ablation corpus480418,695staged4 groups: verse-religious, prose-religious, verse-secular, prose-secular. 120 passages each. For the causal ablation test.

Raw sources

SourceHow muchNotes
Wikipedia theology articles7,508 articlesStarted from 6 seed topics, followed categories 3 levels deep
Project Gutenberg theology books3,827 books40+ topic searches; about 400M tokens total
Project Gutenberg Christian books3,705 books32 topics spanning patristics through modern
Secular sources (for Exp 13)5 worksShakespeare sonnets, Whitman, Dickinson, Darwin, Thoreau

What a record looks like

Each piece of text becomes two files: the input (what we fed the model) and the observation (what we recorded).

Input (the text we feed in)

{"sample_id": "bible_gen_001", "category": "bible",
  "sample_index": 0, "seqlen": 880, "token_ids": [...],
  "text": "In the beginning God created..."}

Output (what the harness records)

{"category": "bible", "sample_index": 0, "seqlen": 880,
  "observation": {"layers": {"0": {"expert_frequencies": [12, 0, 3, ...],
    "gate_weights": [...], "activation_norms": [...]}}}}

The expert_frequencies array has 256 numbers, one per expert, counting how many tokens selected that expert at that layer.

Where to download the data

DatasetWhat's in itLink
Consolidated (45 files)Everything: observations, J-lens data, manifests, analysis, code0xSero/deepseek-v4-flash-reap
Raw observations (11 files)Streaming parquet shards from the observation runs0xSero/deepseek-v4-flash-religious-reap-observations