finepdf-subsampled-embeddings

Public full-text embeddings of storagestuffnobody/finepdf-subsampled. Status: complete. Full-text source rows committed: 8,780,394. Chunk vectors saved: 24,447,135; previously computed first chunks reused: 5,920,000. This dataset is incomplete until status is complete.

The default configuration contains full-text rows once its first checkpoint is published. The separate legacy_prefix configuration retains 5,920,000 original prefix-only rows. These older vectors alone are not full-document aggregates. They are also preserved on the immutable prefix-8k-v1 branch at f06e6ce73a2d89c29c9869010ee63d67e527f4f7.

Each full-text row retains all original text and metadata, every normalized 768-dimensional float32 vector in chunk_embeddings, parallel chunk_metadata, and the final embedding. FinePDF's original train/test membership and Stack's nested file metadata are preserved. Stories retains extracted text, chapter boundaries, original HTML, AO3 metadata, HTTP/WARC headers, URLs and archive offsets from the primary compressed story WARC files.

Consecutive chapters are packed when available; oversized chapters and other documents are packed at the last double-newline boundary that fits. A paragraph too large to fit is split at the token limit. Previously saved first 8K vectors keep their original boundary, as requested. Continuations include up to 256 preceding content tokens from the same chapter. Every model input is at most 8,192 tokens, including the prompt and BOS/EOS. No content tokens are discarded.

The complete title: <title or none> | text: <text> input is tokenized once. Chunks repeat its original prompt-token prefix and slice the canonical content-token stream, avoiding lossy decode/re-encode changes. Token ranges are zero-based and end-exclusive. Character offsets refer to the serialized source text; a byte-fallback token boundary can overlap a character. input_sha256 identifies each exact int32 little-endian model input. Saved first chunks are reused only under the same encoder contract and with the same input length/token construction.

For chunk i, weight = token_end - new_token_start. The aggregate is the L2-normalized sum of normalized chunk vectors multiplied by these new-token counts. Overlap, prompt tokens and special tokens add no weight. Accumulation uses float64, then stores float32. Empty text keeps one prompt-only vector with zero content-token count. embedding_tokens is the total canonical content-token count, embedding_chunks is the number of retained chunks, and embedding_model_tokens includes prompt/special/overlap tokens across all chunks.

Model: google/embeddinggemma-2 at 914f7f89142e33e77833254d9c9b90c3cef7303b. The optimized TileLang path uses FP8 attention, BF16 projections and FP32 accumulation, with masked padding. Chunks cannot attend to each other; aggregation can blur different topics. All chunk vectors are retained for inspection. Numerical agreement has been checked; no downstream retrieval quality evaluation is claimed.

Eight MI325X GPUs process batches of 32 chunks each. Atomic public checkpoints contain at most 10,000 source records (smaller at source/split boundaries). _checkpoints/state.json is the authoritative full-text resume state. Full-text proofs are in _checkpoints/chunked/commits; old files remain available without mixing their schema into the new default configuration. FinePDF and the frozen Stack snapshots run before Stories. Original Stack is fixed at ea897b29c28e2b6dd993987e36af228a3e34986c; Part 2 uses its separately pinned source plan. Later source uploads are excluded. Numerical-reference hashes remain in contract; the actual execution code and scheduling version are recorded separately in checkpoint runtime.

Original content licensing and attribution remain applicable. FinePDF derives from HuggingFaceFW/finepdfs; Stack derives from HuggingFaceCode/stack-v3-train with original per-file license metadata retained. See the source cards for provenance.