code-pii (v3, cross-document coreference)

Span-labeled source files for fine-tuning an LLM / token classifier to detect code-related sensitive data (passwords, API keys, tokens, private keys, emails, and since v3 synthetic person names / usernames) and to anonymize it consistently across files: records are grouped into clusters (real multi-file repos, synthetic bundles, singletons), the same entity recurs across the files of a cluster under one entity_id, and every span carries a gold replacement value. Out of scope: addresses, phone numbers, SSNs.

Built by taking permissively-licensed real files from codeparrot/github-code-clean and injecting format-correct synthetic secrets at positions and in constructs whose distribution is learned from the human-annotated Samsung/CredData corpus, plus labeled hard negatives (placeholders, hashes, UUIDs, env-var references, public keys, example.com emails, placeholder names). ~34% of files are left without any secret as pure negatives. CredData's real labeled lines ship separately as a held-out evaluation set.

Full field-by-field documentation: DATA_DICTIONARY.md. Counts: stats.md.

Contents

path	description
data/train-AAAAA-BBBBB.parquet	generated training records; each file merges 3 base-corpus shards (~225k records, ~550 MB, zstd); clusters are contiguous and never cross files
sample.jsonl	~500 records (whole clusters), same schema, for a quick look
eval_creddata.jsonl (stored gzipped in the project store as eval_creddata.jsonl.gz; upload uncompressed)	67,896 real labeled lines from CredData (+-3 lines of context): 15,753 true secrets, 48,061 annotated false positives, 4,082 ambiguous. Evaluation only.
priors.json	CredData-derived placement statistics that drive the generator
stats.md	label / value_type / context / language counts, sizes, throughput
DATA_DICTIONARY.md	construction narrative, schema, taxonomy, limitations
notebooks/eda.ipynb (+ stats_summary.json)	executed EDA notebook: integrity check, distributions, comparison against CredData priors, conclusions
generate_dataset.py, build_priors.py, validate.py, to_parquet.py, preview_subset.py, make_stats.py, requirements.txt	everything needed to regenerate at any size

Schema

id              string   ghcc-<shard>-<sha1(repo/path)[:16]>
source          string   "codeparrot/github-code-clean"
repo, path      string   origin of the base file
license         string   mit | apache-2.0 | bsd-3-clause | bsd-2-clause | isc | unlicense | cc0-1.0
language        string   30 github-code-clean language names
text            string   full file after injection
spans           list     secret / person spans (may be empty)
hard_negatives  list     look-alike spans labeled not_secret (may be empty)
  span: {start:int64, end:int64, label:str, value_type:str, context:str|null, synthetic:bool, value:str,
         entity_id:str|null, surface_form_type:str|null, replacement:str|null}        # v3 (null in hard_negatives)
cluster_id      string   v3  c-<shard>-<hash>; records of a cluster are contiguous, ordered by doc_index
cluster_kind    string   v3  repo | bundle | singleton
cluster_size    int32    v3  files in the cluster
doc_index       int32    v3  0..cluster_size-1
entities        list     v3  the cluster's entity table, only on doc_index == 0 (null elsewhere)
  entity: {entity_id, entity_type, value_type, canonical, forms{literal,full_name,first_name,last_name,username,handle,email},
           replacement_canonical, replacement_forms{...}, owner_entity_id|null, n_mentions:int32, n_docs:int32}

Coreference: spans with the same entity_id (within a cluster) are the same entity; surface_form_type says which form of it (literal, escaped_literal, or a person form); replacement is the gold anonymized string for that span, derived from one replacement value per entity, so replacing every span by its replacement gives a consistently anonymized cluster. Split train/validation by cluster_id, never by record.

Offsets are 0-based Unicode code-point offsets into text, end-exclusive: text[start:end] == value holds for every span (asserted at generation time and again by validate.py). Convert to bytes with len(text[:start].encode("utf-8")).

label in spans: api_key, password, token, private_key, email, and (v3) person_name, username. In hard_negatives it is always not_secret. Generate with --no-persons for the v2 label set.

value_type (positives): aws_access_key_id, aws_secret_access_key, google_api_key, stripe_secret_key, sendgrid_api_key, twilio_api_key, mailgun_api_key, openai_api_key, anthropic_api_key, azure_client_secret, generic_hex_key, generic_hex_key_upper, generic_base64_secret, generic_alnum_key, generic_lower_key, generic_upper_key, generic_digit_key, generic_hex_dashed_key, github_pat, github_fine_grained_pat, slack_token, slack_webhook, jwt, gitlab_pat, npm_token, discord_bot_token, digitalocean_pat, hf_token, bearer_token_opaque, basic_auth_b64, generic_hex_token, generic_alnum_token, generic_dashed_token, generic_upper_token, password, pem_private_key, email, and (v3, persons) full_name, first_name, last_name, username, handle. value_type (hard negatives): placeholder, example_email, uuid, sha256_hash, sha1_git_sha, md5_hash, env_var_reference, public_key_pem, ssh_public_key, version_or_id, base64_non_secret, placeholder_name (v3).

context (how the secret was rendered): assignment, json_key, yaml_key, env_assignment, comment, url_credential (span = the password inside scheme://user:PASS@host), auth_header (span = the token after Bearer /Basic ), function_arg, pem_block.

Example (abridged; a 2-file cluster, person e000 owns the password e001, which recurs in both files):

{"id":"ghcc-00000-8dd9aed85c603418","cluster_id":"c-00000-cb9f40413a8d75c1","cluster_kind":"repo","cluster_size":2,"doc_index":0,"language":"GO","path":"cmd/db.go",
 "text":"package main\n\n// Copyright (c) 2017 Vasana Sur\nvar dsn = \"postgres://vsur:healthroot@db.internal:5432/app\"\n",
 "spans":[{"start":34,"end":44,"label":"person_name","value_type":"full_name","context":"comment","synthetic":true,"value":"Vasana Sur","entity_id":"c-00000-cb9f40413a8d75c1/e000","surface_form_type":"full_name","replacement":"Lucky Bala"},
          {"start":67,"end":71,"label":"username","value_type":"username","context":"url_credential","synthetic":true,"value":"vsur","entity_id":"c-00000-cb9f40413a8d75c1/e000","surface_form_type":"username","replacement":"lbala"},
          {"start":72,"end":82,"label":"password","value_type":"password","context":"url_credential","synthetic":true,"value":"healthroot","entity_id":"c-00000-cb9f40413a8d75c1/e001","surface_form_type":"literal","replacement":"pooledavid"}],
 "hard_negatives":[],
 "entities":[{"entity_id":"c-00000-cb9f40413a8d75c1/e000","entity_type":"person","value_type":"person","canonical":"Vasana Sur","forms":{"full_name":"Vasana Sur","first_name":"Vasana","last_name":"Sur","username":"vsur","handle":"@vsur","email":"vasana.sur@gmail.com"},"replacement_canonical":"Lucky Bala","replacement_forms":{"full_name":"Lucky Bala","first_name":"Lucky","last_name":"Bala","username":"lbala","handle":"@lbala","email":"lucky.bala@gmail.com"},"owner_entity_id":null,"n_mentions":3,"n_docs":2},
             {"entity_id":"c-00000-cb9f40413a8d75c1/e001","entity_type":"password","value_type":"password","canonical":"healthroot","forms":{"literal":"healthroot"},"replacement_canonical":"pooledavid","replacement_forms":{"literal":"pooledavid"},"owner_entity_id":"c-00000-cb9f40413a8d75c1/e000","n_mentions":3,"n_docs":2}]}
{"id":"ghcc-00000-369c565b6f43c31b","cluster_id":"c-00000-cb9f40413a8d75c1","cluster_kind":"repo","cluster_size":2,"doc_index":1,"language":"GO","path":"pkg/types.go",
 "text":"...\n\tauthor_email: \"vasana.sur@gmail.com\",\n\tvar password = \"healthroot\"\n",
 "spans":[{"...":"email span, entity_id .../e000, surface_form_type email, replacement lucky.bala@gmail.com"},{"...":"password span, entity_id .../e001, replacement pooledavid"}],
 "hard_negatives":[],"entities":null}

How it is built (short)

1. Priors (build_priors.py): parse CredData's meta/*.csv (67,896 rows with line + column offsets, category, GroundTruth) and the rebuilt, CredData-obfuscated files. Collapse ~200 categories to the label set, classify each true secret's syntactic context by regex on the text before the value, collect variable-name vocabularies, value length/charset stats, quote style, extension frequencies and secrets-per-file histogram into priors.json. Emit every labeled line as eval_creddata.jsonl.
2. Base corpus: stream codeparrot/github-code-clean parquet shards over HTTPS (ungated, no click-through), keep rows whose license is MIT/Apache-2.0/BSD/ISC/Unlicense/CC0 (~63%) and 200 B-200 KB.
3. Clustering (v3, generate_dataset.py): per shard, files of the same repo form repo clusters (<=8 files), the rest stay singletons (35%) or are grouped into bundles of 2-8 unrelated files; each cluster gets a seeded entity pool: 0-3 synthetic persons (Faker personas with name / username / handle / email forms) and secrets created on demand, each with a format-preserving gold replacement.
4. Injection: 75% of files get 1-5 secrets (CredData histogram) and 0-3 hard negatives; label, construct and variable name are sampled from the priors; the value's charset class is sampled from CredData's per-label charset distribution and a matching generator is chosen; a secret slot reuses a cluster entity with p=0.7 (p=0.15 repeats one already in the file), otherwise creates a new one, owned by a cluster person with p=0.5 (owned secrets expose the owner's username in user:pass@host, login as, CREATE USER, basic-auth); files in clusters with persons get 0-2 person mentions (comments like @author, Copyright, TODO(user), assignments, JSON/YAML keys); a 31-language syntax table renders the line (assignment, dict/object key, .env/export, comment, URL credential, Authorization header in the language's HTTP idiom, PEM as heredoc/template literal/string concatenation); the line is inserted at a random line boundary with the neighbor's indentation; offsets are asserted.
5. Validation (validate.py): offsets re-checked (0 failures), label tables, regex scanner with gitleaks-style patterns finds 100% of injected spans of 15 pattern-bearing types at the labeled offsets; v3 adds hard checks for cluster contiguity, entity-table consistency (every span's value resolves to its entity's form, every entity has spans, counts match, owner links resolve), replacement validity (same format/charset/length, unique within the cluster, gold anonymized text keeps the scanner hits and leaks no original), and prints 20 random clusters for review.

Statistics

Full v3 release (shards 0-69, seed 2): 5,278,400 records in 2,379,311 clusters (records: 40% bundles, 38% real-repo clusters, 22% singletons; mean 2.22 files/cluster), 3,914,548 entities, of which 1,167,239 / 2,885,220 (40.5%) in multi-file clusters recur in >=2 files; 7,936,370 labeled spans (api_key 37%, password 18%, token 13%, email 12%, private_key 8%, person_name 7%, username 5%), 4,950,684 hard negatives, and 66.2% of files with >=1 span. Size: 47.86 GB JSONL / 11.92 GB Parquet (24 files). The cluster-preserving preview is 1,900,308 records / 856,252 clusters / 4.28 GB Parquet. Detailed tables are in stats.md.

Regenerating the full 70-shard release

pip install -r requirements.txt
# generate exactly the released 70 source shards (downloads on the fly, deletes after use)
python generate_dataset.py --out ./out-v3 --target-bytes 1e12 --max-shards 70 --workers 8 --seed 2
# optional: convert to parquet (zstd, ~2.3x smaller) and aggregate stats
python to_parquet.py ./out-v3 ./out-v3-parquet 8
python make_stats.py ./out-v3 stats.md ./out-v3-parquet
python validate.py --data ./out-v3 --sample-files 2 --write-sample sample.jsonl
# optional: a seeded cluster-preserving subset (e.g. 36%) directly as parquet
python preview_subset.py --src ./out-v3 --out ./preview_parquet --frac 0.36 --seed 42

Measured throughput (v3): ~12 MB/s of JSONL per worker (1.5x slower than v2 because of the entity pool, replacements and the cluster-planning pass; CPU-bound; shard download is ~35 MB/s per stream and overlaps with processing). The measured full 70-shard run took 16.1 min on 4 vCPU including overlapped downloads (49.7 MB/s aggregate), plus Parquet conversion and validation. Flags: --inject-frac (default 0.75), --start-shard/--max-shards to pick disjoint shard ranges for train/validation, --shard-dir --keep-shards to reuse downloaded shards, --min-bytes/--max-bytes; v3: --cluster-cap 8, --singleton-frac 0.35, --p-reuse 0.7, --p-repeat 0.15, --p-owner 0.5, --no-persons (v2 label set), --no-bundles, --no-repo-clusters.

To rebuild priors.json / eval_creddata.jsonl: clone CredData, pip install -r CredData/requirements.txt && python download_data.py --jobs 12 (clones 337 repos, ~14 GB temporary), then python build_priors.py --creddata CredData. To add only the v3 cluster/entity fields to an existing eval file: python build_priors.py --creddata . --annotate-existing eval_creddata.jsonl --out-eval eval_creddata.jsonl.

Using eval_creddata.jsonl

Real code, real annotations, obfuscated secret values (CredData's pipeline replaces every true credential with a format-preserving fake). ground_truth == "T" rows are positives (span available for 14,626 of 15,753), "F" rows are annotated false positives (treat as hard negatives), "X" rows are ambiguous (skip). Its label set includes uuid and other, which the synthetic data treats as negatives/does not emit; map accordingly. Do not train on it. v3 adds cluster_id (= CredData repo) and entity_id by exact value match within the repo, but only 14 of 14,598 distinct secret values recur in more than one file, so this set measures detection; evaluate coreference and anonymization on held-out synthetic clusters (e.g. the last shards of a run) instead.

Caveats

• Injected lines are syntactically plausible but semantically unrelated to the file and placed at random lines; models can learn "out-of-place line" as a cue. Mix with real data (CredData, your own) before trusting precision numbers.
• The base corpus was never scrubbed for secrets or PII: pre-existing real emails/keys in base files are unlabeled (label noise in the negative class).
• Hex key positives vs hash negatives are the same character class and are disambiguated only by variable name, by design.
• Variable-name vocabulary inherits CredData's repo skew (sharedsecret, nonce, salt are over-represented).
• Language mix is code-heavy; YAML/JSON/.env/properties files, where CredData sees most real secrets, are under-represented because github-code-clean has few of them.
• Charset classes are calibrated to CredData (v2) but value lengths within a class and the vendor-prefix mix are generator choices; CredData classes with <1% mass (upper-only, digit+symbol) are not produced.
• Person labels cover injected persons only: real author names/usernames/emails already in the base files are unlabeled (noisier negatives for person_name/username than for secrets).
• bundle clusters are unrelated files; cross-document recurrence is always the literal string (no indirection through env vars / config keys).
• Secrets, emails, names and usernames only; no addresses/phones/SSNs.

Changelog

version	date	change
v1 (PoC)	2026-11-05	24 shards, 1.81M records, 14.4 GB JSONL / 3.84 GB Parquet (one 160 MB file per shard). EDA found value charsets diverged from CredData: api_key 25% plain hex vs 70% real; passwords <1% pure lowercase vs 49% real; tokens 93% mixed-case+symbol vs 40% real.
v2	2026-11-05	Charset recalibration: gen_value now samples a target charset class from priors.json: charset_by_label and picks a generator by `weight x P(class
v3	2026-11-06	Cross-document coreference: records grouped into clusters (repo / bundle / singleton; new cluster_id, cluster_kind, cluster_size, doc_index), per-cluster entity pool with recurring secrets and synthetic persons (new labels person_name, username; new span fields entity_id, surface_form_type, replacement; entity table embedded on doc 0), gold format-preserving replacements for every entity, person-owned secrets (owner_entity_id), placeholder_name hard negatives, cluster-aware validate.py / preview_subset.py / make_stats.py, eval_creddata.jsonl annotated with repo clusters and exact-match entity ids. All v2 fields unchanged.

Licenses

• Base corpus: codeparrot/github-code-clean (dataset compilation Apache-2.0; ungated). Only files whose per-file license is MIT, Apache-2.0, BSD-2/3-Clause, ISC, Unlicense or CC0-1.0 are included and each record carries that license; attribution requirements of those licenses apply to the repo/path listed in each record.
• Seed / eval: Samsung/CredData, Apache-2.0 for the repository and annotations; each embedded source file keeps its own permissive license (see CredData's snapshot.json and per-repo LICENSE handling).
• Synthetic values were produced by the generator and Faker; any resemblance to a real credential is coincidental and all values are inert.
• Generator code in this repo: Apache-2.0.