Datasets:

MannyMorningstar
/

APOC-embedding

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
config: struct<type: string, dimensions: int64, nlist: int64, train_size: int64, seed: int64, train_iteratio (... 384 chars omitted)
  child 0, type: string
  child 1, dimensions: int64
  child 2, nlist: int64
  child 3, train_size: int64
  child 4, seed: int64
  child 5, train_iterations: int64
  child 6, binary_derivation: string
  child 7, source: struct<vectors: struct<dimensions: int64, kind: string, model: struct<max_length: int64, name: strin (... 214 chars omitted)
      child 0, vectors: struct<dimensions: int64, kind: string, model: struct<max_length: int64, name: string, repo: string, (... 53 chars omitted)
          child 0, dimensions: int64
          child 1, kind: string
          child 2, model: struct<max_length: int64, name: string, repo: string, revision: string>
              child 0, max_length: int64
              child 1, name: string
              child 2, repo: string
              child 3, revision: string
          child 3, projection: string
          child 4, scale: int64
      child 1, chunking: struct<overlap: int64, sample_fraction: double, target: int64, tokenizer_sha256: string>
          child 0, overlap: int64
          child 1, sample_fraction: double
          child 2, target: int64
          child 3, tokenizer_sha256: string
      child 2, corpus_digest: string
      child 3, max_chunk_id: int64
  child 8, faiss_version: string
file: string
sha256: string
indexed: int64
last_chunk_id: int64
expected: int64
complete: bool
index_bytes: int64
build_threads: int64
seconds_this_run: double
recommended_search: struct<nprobe: int64, candidates: int64>
  child 0, nprobe: int64
  child 1, candidates: int64
rerank_store: struct<vectors: struct<file: string, sha256: string>, ids: struct<file: string, sha256: string>>
  child 0, vectors: struct<file: string, sha256: string>
      child 0, file: string
      child 1, sha256: string
  child 1, ids: struct<file: string, sha256: string>
      child 0, file: string
      child 1, sha256: string
community_approved: bool
allocated_ram_GB: int64
allocated_vcpu_count: int64
disk_gb: int64
rate_total_hour: double
gpu: string
wrapup_minutes: int64
soft_stop_usd: double
full_run_authorized: bool
requested_unique_name: string
outer_limit_minutes: int64
test: string
created_epoch: double
id: string
total_cap_usd: double
terminate_minutes: int64
disk_rate_hour_estimate: double
hard_kill_usd: double
gpu_rate_hour: double
to
{'test': Value('string'), 'id': Value('string'), 'created_epoch': Value('float64'), 'rate_total_hour': Value('float64'), 'requested_unique_name': Value('string'), 'gpu': Value('string'), 'gpu_rate_hour': Value('float64'), 'disk_gb': Value('int64'), 'disk_rate_hour_estimate': Value('float64'), 'full_run_authorized': Value('bool'), 'community_approved': Value('bool'), 'soft_stop_usd': Value('float64'), 'hard_kill_usd': Value('float64'), 'total_cap_usd': Value('float64'), 'wrapup_minutes': Value('int64'), 'terminate_minutes': Value('int64'), 'outer_limit_minutes': Value('int64'), 'allocated_vcpu_count': Value('int64'), 'allocated_ram_GB': Value('int64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              config: struct<type: string, dimensions: int64, nlist: int64, train_size: int64, seed: int64, train_iteratio (... 384 chars omitted)
                child 0, type: string
                child 1, dimensions: int64
                child 2, nlist: int64
                child 3, train_size: int64
                child 4, seed: int64
                child 5, train_iterations: int64
                child 6, binary_derivation: string
                child 7, source: struct<vectors: struct<dimensions: int64, kind: string, model: struct<max_length: int64, name: strin (... 214 chars omitted)
                    child 0, vectors: struct<dimensions: int64, kind: string, model: struct<max_length: int64, name: string, repo: string, (... 53 chars omitted)
                        child 0, dimensions: int64
                        child 1, kind: string
                        child 2, model: struct<max_length: int64, name: string, repo: string, revision: string>
                            child 0, max_length: int64
                            child 1, name: string
                            child 2, repo: string
                            child 3, revision: string
                        child 3, projection: string
                        child 4, scale: int64
                    child 1, chunking: struct<overlap: int64, sample_fraction: double, target: int64, tokenizer_sha256: string>
                        child 0, overlap: int64
                        child 1, sample_fraction: double
                        child 2, target: int64
                        child 3, tokenizer_sha256: string
                    child 2, corpus_digest: string
                    child 3, max_chunk_id: int64
                child 8, faiss_version: string
              file: string
              sha256: string
              indexed: int64
              last_chunk_id: int64
              expected: int64
              complete: bool
              index_bytes: int64
              build_threads: int64
              seconds_this_run: double
              recommended_search: struct<nprobe: int64, candidates: int64>
                child 0, nprobe: int64
                child 1, candidates: int64
              rerank_store: struct<vectors: struct<file: string, sha256: string>, ids: struct<file: string, sha256: string>>
                child 0, vectors: struct<file: string, sha256: string>
                    child 0, file: string
                    child 1, sha256: string
                child 1, ids: struct<file: string, sha256: string>
                    child 0, file: string
                    child 1, sha256: string
              community_approved: bool
              allocated_ram_GB: int64
              allocated_vcpu_count: int64
              disk_gb: int64
              rate_total_hour: double
              gpu: string
              wrapup_minutes: int64
              soft_stop_usd: double
              full_run_authorized: bool
              requested_unique_name: string
              outer_limit_minutes: int64
              test: string
              created_epoch: double
              id: string
              total_cap_usd: double
              terminate_minutes: int64
              disk_rate_hour_estimate: double
              hard_kill_usd: double
              gpu_rate_hour: double
              to
              {'test': Value('string'), 'id': Value('string'), 'created_epoch': Value('float64'), 'rate_total_hour': Value('float64'), 'requested_unique_name': Value('string'), 'gpu': Value('string'), 'gpu_rate_hour': Value('float64'), 'disk_gb': Value('int64'), 'disk_rate_hour_estimate': Value('float64'), 'full_run_authorized': Value('bool'), 'community_approved': Value('bool'), 'soft_stop_usd': Value('float64'), 'hard_kill_usd': Value('float64'), 'total_cap_usd': Value('float64'), 'wrapup_minutes': Value('int64'), 'terminate_minutes': Value('int64'), 'outer_limit_minutes': Value('int64'), 'allocated_vcpu_count': Value('int64'), 'allocated_ram_GB': Value('int64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

APOC offline knowledge index

Text chunks, embeddings and search indexes for APOC, an offline survival and rebuilding knowledge appliance (medicine, practical skills, engineering, science). The appliance searches this library locally and answers questions with a local model that cites numbered sources. No network is used at runtime.

This dataset is public. It is a working build area, not a single finished release: read the status table before using a package.

What is here

Path	Contents	Status
`combined/apoc-combined-v2/`	Re-embedded vectors for the complete library (24,825,562 chunks): BGE-small on title + section + text, up to 512 tokens, 384-d int8, no projection; binary IVF ANN (16,384 lists) with exact int8 rerank store; Tantivy keyword index (`keyword/tantivy-index.tar`); `evaluation.json`. Text, metadata and FTS5 stay in v1 (same chunk ids)	Current. Use this package. Built 2026-11-11; vector gold@10 72 / 64 / 52% (original / paraphrase / typo questions) vs 50 / 50 / 34% for v1
`combined/apoc-combined-v1/`	The complete library: 24,825,562 chunks (58 non-Wikipedia sources + practical Wikipedia subset) with one FTS5 index (`fts.sqlite.gz`, 13.2 GB, 41.3 GB unpacked), one binary-IVF ANN (16,384 lists) with exact int8 rerank store, float16/text shards, `recall.json` and a checksummed `manifest.json`. Library chunk IDs unchanged; Wikipedia IDs offset by 1,000,000,000	Text, metadata and FTS5 for v2; its 256-d vectors are superseded by v2. Built 2026-11-09
`/` (root `manifest.json`, `chunks-`, `vectors-`, `ids-*`, `fts.sqlite.gz`, `metadata.jsonl.gz`, `ann/`)	940,838 chunks from 4 sources (Wikibooks, Appropedia, LibreTexts Physics, Outdoors Stack Exchange)	Complete; `manifest.json` authenticates every file by SHA-256
`wikipedia-subsets/practical-20261026T161444Z-expanded-final/`	Practical Wikipedia subset: 2,899 articles, 58,317 chunks, 626 checklist topics, 0 unresolved	Current Wikipedia selection; complete and verified
`wikipedia-subsets/practical-20261026T161444Z-expanded/`, `wikipedia-subsets/practical-20261026T075843Z/`	Earlier versions of the subset	Superseded by `-expanded-final`; kept for provenance
`full-runs/full-20261019T155435Z/`	63 shards (44.3M chunks, 36.4M vectors) from the 59-source run, plus two partial recovery checkpoints (`knowledge.db.gz`, ~28-29 GB each)	Incomplete: the run stopped while processing full English Wikipedia, which was later replaced by the practical subset. Shards include partial full-Wikipedia chunks that are outside the agreed selection
`evaluations/combined-qwen3-4b-20261026/`	Exploratory answer test over 24,825,562 chunks (58 non-Wikipedia sources + practical subset) with Qwen3-4B	Six exploratory questions, not a benchmark; manual review found grounding problems
`test4-results/`	Earlier pipeline benchmark results	Historical evidence

combined/apoc-combined-v1/ is the single combined package. Search it with nprobe 1024 and 10,000 candidates (recall@10 vs exact int8 search: 0.888 on 50 real questions, 0.928 on 450 corpus queries). Lower settings are faster but lose recall; see recall.json.

Data contract

v2 (current): same model and revision, same query prefix, max length 512, passage text = title, section (if different) and text, CLS pooling, L2-normalized 384-d, no projection, int8 = rint(clip(x, -1, 1) * 127). ANN codes are int8 >= 0 bits (384). Recommended search nprobe 1024, 10,000 candidates. Never mix v1 and v2 vectors.

v1 (below) and the older packages:

All packages in this dataset share one embedding contract. Query vectors must be produced exactly this way, or results are meaningless:

Encoder: BAAI/bge-small-en-v1.5, revision 5c38ec7c405ec4b44b94cc5a9bb96e735b38267a, max length 256, CLS pooling, L2-normalized 384-dimensional output.
Queries are prefixed with exactly Represent this sentence for searching relevant passages: . Passages have no prefix.
Projection 384 → 256 with a Gaussian matrix seeded 20261017 + 384, divided by sqrt(256), then L2-normalized again. This is a projection, not truncation.
vectors-*.npy: projected normalized float16. ids-*.npy: int64 chunk IDs.
ann/: int8 vectors (rint(clip(x, -1, 1) * 127)) and a FAISS IndexBinaryIVF_HNSW16 over sign bits (int8 >= 0), for candidate generation followed by exact int8 squared-L2 reranking. Binary codes are not the original vectors. Recommended: nprobe 64, 5,000 candidates.
Chunking: target 320 tokens, overlap 48, tokenizer SHA-256 d241a60d5e8f04cc1b2b3e9ef7a4921b27bf526d9f6050ab90f9267a1f9e5c66.
fts.sqlite.gz: gzip of a SQLite FTS5 index (porter unicode61 remove_diacritics 2). Unpack and check unpacked_sha256 in the manifest before use.
chunks-*.jsonl.gz: chunk text and source metadata. metadata.jsonl.gz: source, unit and settings records.

A configuration with max length 512 (used in some local development samples) is a different contract and must not be mixed with these vectors.

Models used by APOC

Both models are mirrored in this dataset under models/ (models/manifest.json has every file's SHA-256), pinned to these upstream revisions:

Role	Upstream	Revision	Mirror
Query encoder	`BAAI/bge-small-en-v1.5` (MIT)	`5c38ec7c405ec4b44b94cc5a9bb96e735b38267a`	`models/encoder/bge-small-en-v1.5/` (graph and tokenizer byte-identical to the corpus build)
Answer model	`unsloth/Qwen3.5-4B-GGUF`, base `Qwen/Qwen3.5-4B` (Apache-2.0)	`e87f176479d0855a907a41277aca2f8ee7a09523`	`models/answer/Qwen3.5-4B-UD-Q4_K_XL.gguf`, 2,912,109,728 bytes, SHA-256 `b252c5610a42ca82d20fe2a12813e9d069eed89292907e26c783eeb0bc961bc7`

The answer model runs on llama.cpp b11711. Disable its thinking by pre-filling an empty <think></think> block (or enable_thinking=false); the /no_think soft switch is unreliable.

Licensing and attribution

This dataset contains text derived from third-party sources. Each source keeps its own license; there is no single license for the whole dataset. Most sources were obtained as Kiwix ZIM archives; the original archive name and SHA-256 are recorded per source in metadata.jsonl.gz and the run reports.

Source families include, among others:

Wikimedia projects (Wikipedia, Wikibooks, Wikivoyage, Wikiversity, Wikisource, Wiktionary): generally CC BY-SA, with attribution via the original article.
Stack Exchange sites: user contributions under CC BY-SA (version depends on post date).
LibreTexts and iFixit: content is often under non-commercial licenses (e.g. CC BY-NC-SA), which restricts commercial reuse of those portions.
Appropedia, Energypedia, WikEM, MDWiki and similar wikis: their own open licenses.
US government publications (e.g. CDC, Ready.gov, USDA, Army publications) are generally public domain; NHS content is under the Open Government Licence; MSF medical guidelines are under MSF's own terms.

A per-source license review has not been completed. If you reuse this data, check the license of each source you use, keep attribution to the original pages, and honor share-alike and non-commercial terms. Nothing here is medical or engineering advice.

Provenance

Built by the APOC preparation pipeline (SQLite FTS5, BGE-small int8 vectors, FAISS binary IVF). Run reports, coverage audits and recovery checkpoints are kept beside each package. Operational run-control files were removed from the current revision on 2026-11-09; earlier revisions remain in this repository's history.

Downloads last month: 1,987

Total file size:

141 GB