tgai commited on
Commit
c048fb9
·
1 Parent(s): 3364fd9

Add dataset card for the-stack-go samples

Browse files
Files changed (1) hide show
  1. README.md +101 -0
README.md ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ pretty_name: the-stack-go samples
4
+ task_categories:
5
+ - text-generation
6
+ tags:
7
+ - code
8
+ - pretraining
9
+ - the-stack-v2
10
+ configs:
11
+ - config_name: 8k
12
+ data_files:
13
+ - split: train
14
+ path: 8k/*.parquet
15
+ ---
16
+
17
+ # the-stack-go-samples
18
+
19
+ Training samples generated from the `s0duku/the-stack-go` ge2 dataset, which is
20
+ a strict `star_events_count > 1` export of The Stack v2.
21
+
22
+ ## Available datasets
23
+
24
+ ### 8k
25
+
26
+ - Directory: `8k/`
27
+ - Maximum context: **8,192 tokens**
28
+ - Shards: **64 Parquet files**
29
+ - Samples: **17,670,028**
30
+ - Total tokens: **100,988,533,457**
31
+ - Source repositories: **3,246,224**
32
+ - Repositories represented in samples: **3,239,349**
33
+ - Parquet size: approximately **58.9 GiB**
34
+ - Tokenizer: `deepseek-ai/DeepSeek-V4-Flash-Base`
35
+ - Tokenizer revision: `8855555deef230a27a21a8d6f294b7b7497759b6`
36
+
37
+ `8k/manifest.json` records the generation configuration and per-shard statistics.
38
+
39
+ ## Schema
40
+
41
+ Each row contains:
42
+
43
+ | Field | Type | Description |
44
+ |---|---|---|
45
+ | `sample_id` | string | Stable sample identifier |
46
+ | `repo_id` | string | Source repository name |
47
+ | `revision_id` | string | Source repository revision |
48
+ | `kind` | string | `dependency` or `target_context` |
49
+ | `text` | large_string | Rendered training text with repository/file boundaries |
50
+ | `files` | list<struct> | File provenance for the rendered content |
51
+ | `token_count` | int32 | Exact token count of `text` |
52
+ | `metadata_json` | large_string | JSON-encoded sample metadata |
53
+
54
+ The nested `files` fields are:
55
+
56
+ ```text
57
+ path, relation, content_policy, role, language, bytes, part, parts
58
+ ```
59
+
60
+ `token_count` is an upper-bounded length, not a fixed length. The maximum is
61
+ 8,192 tokens, while individual samples may be shorter.
62
+
63
+ ## Reading
64
+
65
+ Stream the dataset instead of loading all Parquet rows into memory:
66
+
67
+ ```python
68
+ from datasets import load_dataset
69
+
70
+ ds = load_dataset(
71
+ "s0duku/the-stack-go-samples",
72
+ "8k",
73
+ split="train",
74
+ streaming=True,
75
+ )
76
+ for row in ds.take(3):
77
+ print(row["repo_id"], row["kind"], row["token_count"])
78
+ ```
79
+
80
+ Or read a single local shard with PyArrow:
81
+
82
+ ```python
83
+ import pyarrow.parquet as pq
84
+
85
+ table = pq.read_table("8k/train-00000-of-00064.parquet")
86
+ print(table.schema)
87
+ ```
88
+
89
+ ## Source and licensing
90
+
91
+ The samples are derived from The Stack v2 and preserve repository/file
92
+ provenance. Source licenses and removal requests remain governed by the
93
+ original dataset and upstream content providers. There is no single license
94
+ for all included source code.
95
+
96
+ ## Generation notes
97
+
98
+ Samples are dependency-aware and repository-scoped. Dependency samples present
99
+ files in dependency-first order. Sample text can contain structure markers such
100
+ as `<|begin▁of▁repo▁name|>` and `<|begin▁of▁file|>`; these are tokenizer
101
+ control markers, not part of the original source content.