babylontechnologies commited on
Commit
0177c60
·
verified ·
1 Parent(s): 7103304

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +81 -0
README.md CHANGED
@@ -1,3 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: other
3
  license_name: license.md
 
1
+ ---
2
+ annotations_creators:
3
+ - expert-generated
4
+ language:
5
+ - code
6
+ license: other
7
+ task_categories:
8
+ - text-generation
9
+ task_ids:
10
+ - language-modeling
11
+ tags:
12
+ - code-curation
13
+ - pretraining-filters
14
+ - code-quality
15
+ - dataset-index
16
+ size_categories:
17
+ - n<1K
18
+ ---
19
+
20
+ # 🌌 Ex Nihilo Code 001: Curated High-Signal Repository Index
21
+
22
+ **A high-density, rigorously filtered metadata index designed for AI labs, code LLM pre-training teams, and data curation researchers.**
23
+
24
+ This dataset repository contains the **100-repository public sample** of the Ex Nihilo Code 001 index[cite: 1]. It represents the strict stars-descending head of our broader commercial dataset, isolating high-value, active software repositories while filtering out low-signal noise[cite: 1].
25
+
26
+ ---
27
+
28
+ ## 📊 Dataset Summary & Methodology
29
+
30
+ Training top-tier code models requires more than just scraping GitHub; it requires structural verification of repository quality. This index was constructed using strict baseline metrics to identify repositories with proven community reliance and sustained engineering effort.
31
+
32
+ **This 100-row sample consists of:**
33
+ * **97 Verified-Cut Rows:** Repositories passing our stringent quality gates, including a minimum threshold of 100 merged Pull Requests[cite: 1].
34
+ * **3 Edge-Case Rows:** Repositories falling just below the 100-merged-PR threshold, included for threshold evaluation and boundary testing[cite: 1].
35
+
36
+ *Note: This release is a metadata index and sample pack. It is not a bulk text corpus for end-to-end training.*[cite: 1]
37
+
38
+ ---
39
+
40
+ ## 📐 Metrics Warning: File Size vs. LOC
41
+
42
+ When evaluating repositories for pre-training corpora, byte footprint is a misleading metric. **GitHub `size_kb` does not represent executable Lines of Code (LOC).**[cite: 1]
43
+
44
+ Raw repository size often includes binary assets, extensive commit histories, and serialized data files that do not contribute to code generation capabilities. Please refer to the chart below for the variance between raw disk size and actual estimated LOC within this dataset.
45
+
46
+ ![Size KB vs Estimated LOC](size_kb_vs_estimated_loc.svg)
47
+ *(See `size_kb_vs_estimated_loc.svg` for the full visual breakdown)*[cite: 1]
48
+
49
+ ---
50
+
51
+ ## 🗄️ Data Dictionary
52
+
53
+ Understanding the exact semantics of our filtering columns is critical for data curation teams. Below is the excerpt of the schema applied to this index.
54
+
55
+ *(Merge the exact contents of `DATA_DICTIONARY_EXCERPT.md` here so researchers can view your definitions for licenses, PR thresholds, and star filters)*[cite: 1, 3]
56
+
57
+ ---
58
+
59
+ ## ⚖️ Licensing & Upstream Compliance (Important)
60
+
61
+ **License Designation:** `Other / Custom Index License (See Upstream)`
62
+
63
+ This dataset is an index of metadata and repository references. **This product does not relicense any of the underlying repositories.**[cite: 1]
64
+
65
+ All repository URLs, source code, and commit hashes referenced within this sample retain their original upstream software licenses (e.g., MIT, Apache 2.0, GPL) as assigned by their respective authors. Any research or commercial team utilizing this index to fetch, scrape, or train on the referenced code must independently review and comply with the specific, individual upstream licenses of those repositories.
66
+
67
+ ---
68
+
69
+ ## 🚀 Unlock the Full Commercial Pack
70
+
71
+ This Hugging Face repository contains only the 100-row sample[cite: 1]. For scaled evaluations, complete pre-training pipelines, and comprehensive metadata, the **Ex Nihilo Code 001 Paid Listing Pack** is available for production use.
72
+
73
+ **The Full Pack Includes:**
74
+ * **`exnihilo_code_001_full.csv`**: The complete 339-row baseline index[cite: 1].
75
+ * **`exnihilo_code_001_verified.csv`**: The 319-row verified cut (all repos passing the >100 merged PR threshold)[cite: 1].
76
+ * **`exnihilo_code_001_premium.csv`**: The exclusive 205-row ultra-high-signal cut[cite: 1].
77
+ * **Frozen Harvest Archive (`exnihilo_code_001_harvest.zip`)**: 1,020 `files.json` and 1,020 `metadata.json` files containing deep structural data for the repositories[cite: 1].
78
+ * **Complete `DATA_DICTIONARY.md`** and **Stats JSON**[cite: 1].
79
+
80
+ **Email nehemiahsturdivant@babylontechnolgoies for the full dataset**
81
+
82
  ---
83
  license: other
84
  license_name: license.md