Ayush-Singh/cwv-strategy-mining-dataset

CWV Strategy Mining Dataset

Real, merged GitHub PRs mined for measured Core Web Vitals (CWV) techniques, plus every intermediate pipeline artifact through final generated playbook candidates. Produced by cwv-playbook-miner, scanning the full public GH Archive event stream. Every decision in the pipeline runs on real PR text (title, body, diff, comments, reviews) — never a short phrase, never a bare similarity threshold. See the repo's SESSION_NOTES.md for the full chronological design log: what was found, why each fix was made, and what it was verified against.

Why

Automated Lighthouse/bundle-size bots are the obvious place to find before/after performance numbers on a PR, but they're a small fraction of real performance-motivated PRs — most authors and reviewers never run one. This dataset combines that narrow, precisely-measured bot signal with a much larger pool of PRs where a human reviewer (or another automated tool) mentioned performance in plain text, with no bot-parsed number required.

Contents

Folder	Files	What it is
source/	perf_improvement.jsonl, perf_decrease.jsonl, perf_flagged.jsonl	Raw mined PR records — repo, PR number, signal type, bot-parsed metric (when available), title, body, diff patches, and every issue comment/review/inline review comment backfilled via GitHub GraphQL
patterns/	extractions.jsonl	Per-PR technique extraction: technique, mechanism, affected_resource, render_phase, description, inferred direction (for human-flagged PRs) — only for PRs that passed the CWV-relevance gate
aggregation/	routing.jsonl, novel_clusters.jsonl, enrichments.jsonl, playbook_facts.jsonl	Routing decisions (existing playbook match vs. novel), coherence-verified novel-technique clusters, diversity-weighted existing-playbook evidence, and the same fact-shape extracted from the 20 curated playbooks
playbooks/new_playbooks/	9 .md files	Full new {issue_type}.md candidates — front matter includes source_prs, draft → critic → grounding-checked → AEM-fidelity-checked
playbooks/enriched/	17 .enrichment.md files	New approach/anti-pattern subsection(s) for an existing playbook — no front matter (spliced body content), grounding noted via a > **Source PRs** line instead

source/*.jsonl schema

{
  "id": "repo#pr_number",
  "repo": "owner/repo",
  "pr_number": 123,
  "signal_type": "perf_improvement | perf_decrease | perf_flagged",
  "metric_key": "performance | lcp_ms | cls | ... | null",
  "before": 70.0, "after": 88.0, "delta": 18.0,
  "title": "PR title",
  "pr_body_markdown": "...",
  "changed_files": [{"filename": "...", "patch": "..."}],
  "pr_comments": [{"kind": "issue_comment|review|review_comment", "author": "...", "body": "...", "created_at": "...", "state": "...", "path": "..."}],
  "text_enriched": true, "text_truncated": false,
  "merged_at": "2025-01-01T00:00:00Z",
  "human_signal_text": "the flagging review/comment text (perf_flagged only)"
}

title/pr_body_markdown/pr_comments are backfilled via GitHub's GraphQL API (enrich-pr-text + refetch-truncated) — GH Archive's free event stream never carries them. perf_flagged records have metric_key/before/after/delta all null (no bot template matched them); human_signal_text carries the comment that triggered discovery.

patterns/extractions.jsonl schema

{
  "record_id": "repo#pr_number", "drop": false,
  "relevance_reasoning": "...",
  "technique": "...", "mechanism": "...",
  "affected_resource": "image|font|javascript|css|network|dom|server-response|third-party-script|media",
  "render_phase": "pre-paint|post-paint|interaction|build-time",
  "description": "...", "direction": "positive|negative|unclear|null"
}

drop=true means the record failed the stage-3 relevance gate (a dedicated yes/no: would this PR's actual code change move LCP, CLS, INP, or a closely related metric) — every other field is empty for those rows.

Numbers

Full run over the 5-year backfilled source corpus (2021–2026):

Source records	14,965
Enriched with real title/body/comments	14,919 (99.7%)
CWV-motivated (passed the relevance gate)	4,163 (27.8%)
Routed to an existing playbook	1,861
Routed as novel	2,302
Raw HDBSCAN clusters	75
Confirmed coherent	20
Final novel techniques	9
Existing playbooks enriched	17
Generated playbook files	26

Collection method

1. Source (GH Archive scan): merged PRs matching a bot before/after- report template, or a non-bot human review/review-comment matching a broad performance vocabulary (no parsed delta for the latter — direction gets judged downstream from real evidence, never assumed from the marker match).
2. Enrich: backfill title/body/comments/reviews via GitHub GraphQL, batched via query aliases; a second paginated pass closes any gap from a PR that had more comments/reviews than the first pass's page cap.
3. Extract: a dedicated yes/no judges CWV relevance first — grounded in specific metrics (LCP/CLS/INP/TTFB/FCP/TBT/bundle size/request count), with forced reasoning before the verdict — then, only for what passes, a second call extracts objective technique facts.
4. Route: embedding similarity narrows each PR to its top-3 candidate existing playbooks; an LLM verifies against their full text before deciding existing vs. novel — never a bare threshold.
5. Cluster: HDBSCAN groups the novel pool, then a dedicated coherence- verification call reads full evidence per candidate cluster and rejects anything that isn't genuinely one technique before it's ever labeled.
6. Generate: draft → critic → grounding check → AEM-fidelity check (rewrites any code example that isn't genuinely native to its claimed AEM flavor, e.g. a carried-over React example), with diversity-weighted evidence selection so one repo can't dominate a technique's evidence.

Experiment specifications

Models (Azure OpenAI, openai-compatible backend): chat completion gpt-5.4-mini for every LLM stage (relevance, extraction, routing verify, coherence, labeling, draft/critic/grounding/AEM-fidelity); text embedding text-embedding-3-small for routing's pre-filter and clustering's HDBSCAN input. Both verified as actually deployed and callable on the Azure Foundry resource before the run — gpt-5 (the .env deployment name) and text-embedding-ada-002 are not callable despite being listed in the model catalog.

Call parameters: temperature=0.0 for structured/JSON calls (relevance, extraction, routing verify, coherence, labeling), temperature=0.2 for free-text generation calls (draft, critic, grounding check, AEM-fidelity check). Transient network/5xx/429 errors retry up to 4 times with exponential backoff; a real 4xx is never retried.

Batching: BATCH_SIZE=10 records/call for extraction, VERIFY_BATCH_ SIZE=6 for routing verify, BATCH_SIZE=40 PRs/request for GitHub GraphQL enrichment, CLUSTER_CALL_BATCH_SIZE=4 clusters/call for coherence-verification and labeling (an unbatched 72-cluster call measured ~2M characters / ~508K tokens and hit an HTTP 400).

Routing: embedding pre-filter narrows each PR to its top TOP_K=3 candidate existing playbooks by cosine similarity before an LLM verifies against their full text.

Clustering: sklearn.cluster.HDBSCAN(min_cluster_size=4, min_samples=2) on L2-normalized embeddings; a cluster also needs distinct_repo_count >= 2 to reach the coherence-verification call.

Timeout: 180s per LLM call.

License

MIT for this dataset's structure/derived content. Source PR text (titles, bodies, patches, review comments) is quoted from public GitHub repositories under their own individual licenses; this dataset does not relicense that content.