Add expanded GitHub ML seed and update dataset card

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+data/observations/2026/09/24/expanded-seed-20260924.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -16,9 +16,11 @@ A broad, evolving index of public GitHub repositories that may present a machine
 
 Each row is a **discovery candidate**, not a verified claim of novelty, correctness, reproducibility, or scientific quality. The collection favors recall and includes borderline candidates. Repository metadata and labels can be incomplete or stale. The registry is intended to support discovery and reduce duplicated Modelome research; it is not a comprehensive census of ML work.
 
+The display name is **GitHub ML**; `gh-ml` is the dataset and code repository slug.
+
 ## Data files
 
-The collector appends machine-readable JSON Lines observations by UTC collection date, with run coverage and a resumable checkpoint, using this layout:
+The collector appends machine-readable JSON Lines observations by UTC publication date, with run coverage and resumable checkpoints, using this layout:
 
 ```text
 README.md
@@ -28,7 +30,7 @@ state/checkpoint.json
 state/backfill.json
 ```
 
-Each JSONL row is one observation, not a unique repository across the full history. Run observations are deduplicated by `github_id` within that run. To create a current snapshot, combine observation files, group by numeric `github_id`, and retain the row with the greatest `observed_at`; `github_id` remains stable if the repository is renamed. Coverage JSON records per-query and date-partition outcomes; `state/checkpoint.json` stores daily collection progress and `state/backfill.json` stores independent historical backfill progress. The collector writes these files; they are not maintained by hand. See [Hugging Face repository structure](https://huggingface.co/docs/datasets/main/repository_structure) and [dataset cards](https://huggingface.co/docs/hub/en/datasets-cards).
+Each JSONL row is one observation, not a unique repository across the full history. A run deduplicates matches by `github_id`, but later runs append new snapshots. To create a current view, combine all observation files, group by numeric `github_id`, and retain the row with the greatest `observed_at`; `github_id` remains stable if the repository is renamed. Keep the full history when change over time matters. Coverage JSON records the query or query/date partitions attempted, result counts, pages scanned, outcome status, and any coverage gap. `state/checkpoint.json` stores daily collection progress and `state/backfill.json` stores independent historical backfill progress. Successful empty sweeps still add coverage and update the checkpoint, but have no JSONL file. Run records and coverage are append-only; checkpoint files and the dataset card are updated as collection continues. See [Hugging Face repository structure](https://huggingface.co/docs/datasets/main/repository_structure) and [dataset cards](https://huggingface.co/docs/hub/en/datasets-cards).
 
 ## Fields
 
@@ -53,29 +55,48 @@ Labels are open vocabulary, multi-label, and subject to change. They can describ
 
 Daily collection searches for recently pushed candidates, records the queries that found each repository, and appends observations. Repository searches include `fork:true` unless a configured query already specifies a `fork:` qualifier, so forks and reproductions can be discovered; `fork` marks forked projects. Within a run, matches merge by numeric `github_id`, never by mutable `name`. Across runs, the observation history is append-only; to materialize a current view, group by `github_id` and select the greatest `observed_at`. Historical backfill searches creation dates beginning at 2008-01-01 by default. It partitions by date and recursively splits dense intervals when GitHub approaches its per-search result ceiling; coverage records each query and date partition. It resumes from the separate `state/backfill.json` checkpoint, while daily progress uses `state/checkpoint.json`. An older backfill observation does not displace a newer snapshot in a current view.
 
-The source is GitHub's public repository metadata and Search API. GitHub Search caps each query at 1,000 returned results and at 4,000 repositories searched, and is subject to request limits, timeouts, incomplete responses, and indexing gaps. Broad query coverage and backfills improve recall but cannot guarantee exhaustiveness. Results can include false positives, and not all novel ML work is hosted on GitHub or discoverable by the configured queries. See the [official Search API documentation](https://docs.github.com/en/rest/search/search).
+The source is GitHub's public repository metadata and Search API. GitHub Search caps each query at 1,000 returned results and at 4,000 repositories searched, and is subject to request limits, timeouts, incomplete responses, and indexing gaps. A coverage `status` of `capped` or `incomplete`, or `coverage_gap: true`, flags known gaps; check `coverage_gap_reason` and the other per-query fields. Broad query coverage and backfills improve recall but cannot guarantee exhaustiveness. Results can include false positives, and not all novel ML work is hosted on GitHub or discoverable by the configured queries. See the [official Search API documentation](https://docs.github.com/en/rest/search/search).
 
 ## Access
 
-The configured destination is [`modelomics/gh-ml`](https://huggingface.co/datasets/modelomics/gh-ml). The dataset has not yet been published; this card describes the intended repository contents.
+This dataset is maintained at [`modelomics/gh-ml`](https://huggingface.co/datasets/modelomics/gh-ml). The collection workflow appends new observations and coverage records to this repository over time. Publishing requires an `HF_TOKEN` with write permission to the dataset repository; the GitHub Actions workflow checks for the secret and exits before collection if it is missing. Actions supplies `GITHUB_TOKEN` for GitHub Search. To recover a scheduled run, use **Run workflow** in Actions: each successful run commits its cursor with the observations and coverage, while a failed search leaves the prior Hub checkpoint available for retry.
+
+For local publishing, set `HF_TOKEN` in the environment or sign in with `uv run hf auth login`; the collector reads the saved Hugging Face CLI token. GitHub authentication can be provided through `GITHUB_TOKEN` or `gh auth login` (the collector reads `gh auth token`). The HF token must have write permission on `modelomics/gh-ml`. For local recovery, rerun using the same `--output-dir` so the local cursor is reused; the default is `~/.local/share/modelomics-gh-ml/runs`. Use `--no-publish` for local-only collection, which writes run files without requiring Hugging Face credentials.
 
-Observation files are JSON Lines stored under date-partitioned Hub paths. To load them with 🤗 Datasets, first download the repository files locally, then pass the local file paths:
+Observation files are JSON Lines stored under date-partitioned Hub paths. To load them with 🤗 Datasets, first download the repository files locally, then pass all observation paths. `datasets` is an additional package for this example (`uv add datasets` in a local environment, or install it separately); it is not needed by the collector.
 
 ```python
 from pathlib import Path
+import json
 from huggingface_hub import snapshot_download
 from datasets import load_dataset
 
 root = Path(snapshot_download(
     repo_id="modelomics/gh-ml",
     repo_type="dataset",
-    allow_patterns=["data/observations/**/*.jsonl"],
+    allow_patterns=["data/observations/**/*.jsonl", "coverage/*.json"],
 ))
 files = sorted(str(path) for path in root.glob("data/observations/**/*.jsonl"))
+if not files:
+    raise RuntimeError("The dataset has no observation files yet")
 observations = load_dataset("json", data_files=files, split="train")
+
+# Current snapshot: one latest row for each stable GitHub repository ID.
+latest = {}
+for row in observations:
+    previous = latest.get(row["github_id"])
+    if previous is None or row["observed_at"] > previous["observed_at"]:
+        latest[row["github_id"]] = row
+current = list(latest.values())
+
+# Coverage is per run; keep it separate from repository rows.
+coverage = [
+    json.loads(path.read_text())
+    for path in sorted((root / "coverage").glob("*.json"))
+]
 ```
 
-For a current repository view, group observations by `github_id` and retain the row with the greatest `observed_at`. Coverage and checkpoint JSON files are operational metadata, not rows in the repository table.
+Coverage is per run; it is not a list of repositories. Daily coverage rows describe pushed-date search windows; backfill rows describe created-date partitions. A `complete_sweep` of `false` usually means the request budget left a cursor to resume. Within each run's `queries`, review `status`, `coverage_gap`, `coverage_gap_reason`, `incomplete_results`, and `search_limit_reached`. Coverage and checkpoint JSON are operational metadata, not rows in the repository table. Hub checkpoints let scheduled or manually dispatched Actions runs resume; locally, reuse the same `--output-dir` to resume local state.
 
 ## License and attribution
 

data/observations/2026/09/24/expanded-seed-20260924.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:896fdd05e6dc76850bcb7e3f6434b400e734abccbcddba92dedba8066f459314
+size 28359666

manifests/expanded-seed-20260924.json

@@ -0,0 +1,17 @@
+{
+  "coverage_file": "coverage-expanded-seed-20260924.json",
+  "failed_query_count": 0,
+  "manifest_file": "manifest-expanded-seed-20260924.json",
+  "observations_file": "observations-expanded-seed-20260924.jsonl",
+  "published": true,
+  "readme_file": "README.md",
+  "repo_id": "modelomics/gh-ml",
+  "run_id": "expanded-seed-20260924",
+  "sample_scope": "one first page of up to 100 repositories per current catalog query, including forks",
+  "selected_query_count": 412,
+  "started_at": "2026-09-24T10:49:00Z",
+  "successful_query_count": 412,
+  "sum_of_query_total_counts": 37709099,
+  "target_revision": "refs/pr/1",
+  "unique_repository_count": 34848
+}