final_report.md

# Final Report — K-12 Math & Coding Dataset

A 1B+ token, K-12-focused math corpus and a 1B+ token educational coding
corpus, built per `pipeline.md`'s 6-phase spec. Both categories comfortably
exceed the 1B-token target after cleaning, deduplication, quality filtering,
and budget-controlled final sampling.

## 1. Total tokens per category (final deliverable)

| Category | Final Tokens | Final Rows | Target | Margin |
|----------|-------------:|-----------:|-------:|-------:|
| Math     | 1,004,991,667 | 676,018 | 1,000,000,000 | +0.5% |
| Code     | 1,020,505,114 | 937,248 | 1,000,000,000 | +2.1% |

Deliverables: `data/final/math/final.jsonl`, `data/final/code/final.jsonl`.
Schema per record: `{text, source, category, quality_score, token_count,
language, metadata}`. `quality_score` is populated only for finemath4plus
(the one source with an upstream classifier score) — `null` elsewhere, since
no equivalent score exists for structured benchmarks or code sources.
`language` is the natural-language code (`"en"` throughout, all sources
verified predominantly English); for code sources the programming language
is preserved under `metadata.programming_language` instead.

## 2. Per-source breakdown (final corpus)

### Math

| Source | Mode | Rows | Tokens | % of Category |
|--------|------|-----:|-------:|---------------:|
| finemath4plus | sampled to 700M | 508,357 | 700,003,574 | 69.7% |
| openwebmath | sampled to 300M | 146,473 | 300,001,270 | 29.9% |
| gsm8k | all | 8,792 | 1,417,863 | 0.14% |
| hendrycks_math | all | 12,396 | 3,568,960 | 0.36% |

### Code

| Source | Mode | Rows | Tokens | % of Category |
|--------|------|-----:|-------:|---------------:|
| githubcode_python | all (under budget) | 236,531 | 287,901,844 | 28.2% |
| githubcode_javascript | sampled to 300M | 333,786 | 300,000,630 | 29.4% |
| githubcode_java | sampled to 240M | 236,242 | 240,000,466 | 23.5% |
| githubcode_html | sampled to 90M | 51,764 | 90,006,022 | 8.8% |
| githubcode_css | sampled to 60M | 50,152 | 60,000,074 | 5.9% |
| freecodecamp | all | 10,258 | 5,175,953 | 0.5% |
| exercism | all | 422 | 275,072 | 0.03% |
| apps | all | 4,959 | 16,080,052 | 1.6% |
| codecontests | all | 13,134 | 21,065,001 | 2.1% |

`githubcode_python` came in under its original 400M config target (Step 4.1's
`invalid_syntax` filter dropped 41,130 Python-2-syntax files — a deliberate,
verified-beneficial filter, not a bug) and was included in full rather than
backfilled by loosening quality thresholds. The shortfall was made up by
increasing javascript/java/html/css's sampled targets above pipeline.md's
original example percentages, since those sources had ample surplus.

## 3. Pipeline stage summary — token counts

| Category | Raw | After Cleaning | After Dedup | After Filtering | Final |
|----------|----:|----------------:|------------:|-----------------:|------:|
| Math | 3,248,617,940 | 3,231,640,208 | 3,215,245,918 | ~3,059,148,111 | 1,004,991,667 |
| Code | 4,589,566,182 | 2,289,366,123 | 2,235,414,819 | ~2,109,766,319 | 1,020,505,114 |

(Code's raw→cleaned drop is large because Step 1.4's `githubcode` source was
deliberately over-downloaded well beyond the final target, per pipeline.md's
"download ~10x headroom" guidance, then filtered down at every stage.)

## 4. Cleaning pipeline summary (Phase 2)

5 rule groups applied across 4 cleaning scripts (`clean_web_math.py`,
`clean_structured_math.py`, `clean_code_stack.py`, `clean_code_educational.py`):
- **Group A** (web-crawled math): boilerplate/nav stripping, spam/warez
  pattern removal, encoding normalization (ftfy), PII redaction.
- **Group B** (structured math): light validation only (already-clean
  benchmark data).
- **Group C** (GitHub code): vendored-code exclusion, autogenerated-file
  exclusion (javadoc-specific + generic), minified-code exclusion, excessive
  encoded-content exclusion, trivial-test exclusion.
- **Group D** (educational code): freeCodeCamp non-coding block exclusion
  (108 language-course superblocks removed via `freecodecamp_excluded_blocks.json`).
- **Group E** (PII): email/phone/IPv4/SSN/API-key redaction with extensive
  context-aware false-positive suppression (math arithmetic, version
  strings, ISBNs, DOIs, test262 spec identifiers — see `config/cleaning_rules.md`).

22 real bugs were found and fixed during Phase 2/4 via direct spot-checking
of actual output (not just exit-code/summary-count trust), documented in full
chronological detail in `progress.md`'s Decisions Log. The single most
consequential: a math-density/language heuristic was wrongly rejecting
excellent K-12 content (41 official AMC 8 competition problems, geometry
"kite/rhombus" problems) because it lacked explicit math vocabulary or used
non-ASCII notation (Greek letters) — fixed by trusting FineMath's own
upstream quality classifier for that source instead of re-applying a cruder
heuristic on top of it.

## 5. Deduplication statistics (Phase 3)

**Exact (SHA-256) dedup:**
| Category | Rows In | Rows Out | Duplicates Removed |
|----------|--------:|---------:|--------------------:|
| Math | 2,181,776 | 2,172,722 | 9,054 (0.4%) |
| Code | 1,974,655 | 1,973,991 | 664 (0.03%) |

**Fuzzy (MinHash LSH, Jaccard 0.8 math / 0.85 code) intra-source dedup:**
| Source | Kept | Removed |
|--------|-----:|--------:|
| finemath4plus | 1,878,340 | 5,128 |
| openwebmath | 267,941 | 21 |
| gsm8k | 8,792 | 0 |
| hendrycks_math | 12,484 | 16 |
| githubcode_python | 289,598 | 3,216 |
| githubcode_javascript | 410,580 | 5,542 |
| githubcode_java | 805,252 | 18,490 |
| githubcode_html | 327,796 | 18,692 |
| githubcode_css | 64,347 | 1,545 |
| freecodecamp | 10,262 | 113 |
| exercism | 422 | 2 |
| apps / codecontests | 5,000 / 13,134 | 0 / 0 |

**Fuzzy inter-source dedup:**
| Category | Total Seen | Total Kept | Removed |
|----------|-----------:|-----------:|--------:|
| Math | 2,167,557 | 2,166,099 | 1,458 |
| Code | 1,926,391 | 1,926,161 | 230 |

A performance bug was found and fixed during this phase: per-shingle
`MinHash.update()` cost scaled with document length (measured 13.6ms/doc,
would have taken ~6+ hours on the largest sources) — fixed by capping
shingles fed into MinHash to 150/doc via deterministic even-spacing
(not random sampling, to stay reproducible), achieving a 4.6x speedup with
`num_perm=128` kept unchanged per spec.

## 6. Quality filter statistics (Phase 4, Step 4.1)

**Math** (`filter_math.py`) drop reasons:
| Source | Records In | Records Out | Top Drop Reasons |
|--------|-----------:|-------------:|-------------------|
| finemath4plus | 1,878,340 | 1,875,104 | reddit_comment_dump (1,108), length_out_of_bounds (1,716), repetition (401), encoding_quality (11) |
| openwebmath | 266,495 | 232,018 | low_math_density (29,603), length_out_of_bounds (3,047), non_english (1,680), repetition (83), reddit_comment_dump (62), encoding_quality (2) |
| gsm8k | 8,792 | 8,792 | none |
| hendrycks_math | 12,472 | 12,396 | length_out_of_bounds (76) |

**Code** (`filter_code.py`) drop reasons:
| Source | Records In | Records Out | Top Drop Reasons |
|--------|-----------:|-------------:|-------------------|
| githubcode_python | 289,598 | 236,531 | invalid_syntax (41,130), comment_ratio_out_of_bounds (10,198), poor_identifier_names (905), length_out_of_bounds (834) |
| githubcode_javascript | 410,571 | 400,453 | comment_ratio_out_of_bounds (6,703), poor_identifier_names (1,586), length_out_of_bounds (1,829) |
| githubcode_java | 805,105 | 771,600 | comment_ratio_out_of_bounds (30,344), length_out_of_bounds (2,409), poor_identifier_names (752) |
| githubcode_html | 327,765 | 324,900 | poor_identifier_names (943), length_out_of_bounds (1,002), comment_ratio_out_of_bounds (920) |
| githubcode_css | 64,304 | 62,405 | comment_ratio_out_of_bounds (1,367), length_out_of_bounds (493), poor_identifier_names (39) |
| freecodecamp / exercism / apps / codecontests | small, near-zero drops | | length_out_of_bounds only |

`invalid_syntax` for Python was investigated directly (14.2% of raw files) and
confirmed to be genuine Python 2 syntax (`print x`, `except X, e:`) — kept as
a deliberate filter since excluding legacy syntax benefits a modern coding
corpus. `comment_ratio` originally had a serious formula bug (a
double-normalized calculation that returned meaningless values, plus a
DOTALL regex bug that let one `#` comment "match" to end-of-file) — both
fixed and the check simplified to a single upper bound (0.8) uniform across
all 5 languages, after measuring that ~54% of real JavaScript files fall
below any reasonable lower bound despite being genuinely good code.

Step 4.2 (optional GPU classifier) was explicitly skipped per pipeline.md's
"optional" designation — not requested by the user, and unnecessary given
ample raw token surplus confirmed at every stage. Step 4.3 was effectively
completed by Step 4.1's scripts.

## 7. Audit results summary (Step 4.4 + Step 5.4)

**Step 4.4 (pre-assembly filtered-data audit, sampled from `data/filtered/`):**
- Code: 39/100 read — ~90% good, ~10% borderline (Sphinx docs, license
  pages), 0% bad. Thresholds met.
- Math: 40/100 read — ~80% good, ~12.5% borderline, 7.5% bad. Thresholds
  met. Found and fixed the Reddit comment-dump content-appropriateness
  issue (`is_reddit_comment_dump`, added despite low 0.065%/0.030% volume
  on appropriateness grounds, not quality-volume grounds).

**Step 5.4 (final post-assembly audit, sampled from `data/final/`):**
- Code: 46/200 read — 93.5% good, 4.3% borderline, 2.2% bad. Thresholds met.
- Math: 40/200 read — 72.5% good, 12.5% borderline, **15% bad (above the
  10% threshold)**. Six distinct low-frequency patterns with no shared
  cheap detection signature (AI-spam filler, off-topic forum dumps,
  content-aggregator splicing, thin SEO templates, a scraping artifact, and
  one essay-mill/homework-cheating-service ad). Flagged to the user with
  three remediation options; **user decision: accept as a documented
  residual limitation and proceed to delivery**, given the corpus is still
  ~85% good+borderline and further tightening risks reintroducing the
  AMC-8 false-positive problem. Full detail in `reports/final_audit.md`.

Full detail for both audit rounds: `reports/quality_audit.md` (Step 4.4),
`reports/final_audit.md` (Step 5.4).

## 8. Known limitations and caveats

- **Math corpus bad-rate (15% at final n=40)**: see above — accepted by the
  user as a residual limitation. Content is genuinely low-quality/off-topic
  in these cases, not corrupted or unsafe at scale, but a downstream
  consumer doing further quality-sensitive work may want to apply additional
  filtering for AI-generated SEO filler, off-topic forum dumps, and
  commercial/service advertisements.
- **AMPS and OpenStax** (pipeline.md's example math sources) were never
  sourced — AMPS has no confirmed public HuggingFace dataset, and OpenStax
  was not investigated given time constraints and ample surplus from the
  sources actually used.
- **The Stack v2** (pipeline.md's primary code source) was gated/inaccessible
  (403 GatedRepoError, confirmed 2026-08-22) and was replaced with
  `codeparrot/github-code-clean` per user decision — a non-gated substitute
  with a different (row-level, not pre-sharded) structure.
- **freeCodeCamp + Exercism combined (~5.5M final tokens)** fell well short
  of pipeline.md's example ~150M "tutorial/educational" allocation — accepted
  by the user, made up by drawing more from the primary GitHub code source
  instead.
- **HTML quality**: flagged early (Step 1.6) as having real quality issues
  (OCR garbage, non-javadoc auto-generated docs slipping through). Kept per
  user decision, relying on Phase 4's heuristic filters — final audit found
  the residual auto-generated-doc gap persists at low frequency (Sphinx,
  Clover, Doxygen artifacts), consistent across both audit rounds.
- **quality_score is null for all non-finemath4plus sources** — no
  equivalent upstream classifier score exists for those sources; this was a
  deliberate choice over fabricating a placeholder value.
- **Duplicate-block detector (built, not deployed)**: a generic
  internal-duplication heuristic was built and tested during the Step 4.4
  math audit (catching 1.5% of finemath4plus) but rejected after inspection
  showed most matches were legitimate forum-quoting structure, not junk
  duplication — documented as a rejected approach rather than shipped.

## 9. Pipeline runtime and process notes

The pipeline ran across 2026-08-22 through 2026-08-24. 22+ real bugs were
found and fixed via direct spot-checking of actual output at every stage
(never trusting exit codes or summary counts alone) — full chronological
detail in `progress.md`. One data-integrity incident occurred late in
Phase 5: two duplicate `filter_math.py --source finemath4plus` processes
were found running concurrently, both writing to the same output file in
`"w"` mode (a race condition from an unreconciled duplicate background
launch), corrupting the file mid-write. Caught during Phase 5 token-summing
(the file's row count changed between two consecutive reads), both
processes were killed, the corrupted output deleted, and a single clean
instance rerun and fully verified before any downstream work relied on it.