aluha501 commited on
Commit
df7e404
·
verified ·
1 Parent(s): 3155059

docs: clarify v0.7 corpus and metadata semantics

Browse files

Mark corpus_rag as the primary content corpus and distinguish interaction metadata from factual labels.

Files changed (1) hide show
  1. v0.7.0-multisource-merged/README.md +22 -13
v0.7.0-multisource-merged/README.md CHANGED
@@ -32,25 +32,34 @@ configs:
32
  path: views/issue_resolution_multisource_merged_v0.1/validation.parquet
33
  ---
34
 
35
- # Community QA Lab v0.7.0 multisource merged — local-only
36
-
37
- This is a local Hugging Face-shaped preprocessing bundle. It combines
38
- the immutable v0.6 Stack Exchange derivative with the v0.7 Stack
39
- Exchange expansion, GitHub Discussions, GitLab retrieval rows and
40
- GitHub/GitLab lifecycle evidence. It has not been uploaded or pushed.
 
 
 
 
 
41
 
42
  ## Merged views
43
 
44
- - preference_dpo_multisource_merged_v0.1: 7453 directed pairs.
45
- Only accepted/marked-answer sources participate; chosen is a weak
46
- community/author interaction label, not factual truth.
47
  - reranking_listwise_multisource_merged_v0.1: 4550 groups.
48
  Candidate metadata is retained for evaluation but excluded from the
49
  model projection.
50
  - corpus_rag_multisource_merged_v0.1: 39215 licensed documents.
51
  document_text contains question context and post text only.
52
- - issue_resolution_multisource_merged_v0.1: 1823 observed
53
- issue/PR lifecycle records; these are not preference targets.
 
 
 
 
54
 
55
  ## Split counts
56
 
@@ -71,8 +80,8 @@ GitHub/GitLab lifecycle evidence. It has not been uploaded or pushed.
71
  no public-text rows.
72
  - Accepted, vote/reaction, timestamps, URLs and provenance fields
73
  remain outside model-input projections.
74
- - This is an experimental retrieval/ranking corpus, not a
75
- contamination-free factuality benchmark.
76
 
77
  The exact source manifest hashes, split adjustments, deduplication
78
  counts and Parquet hashes are in manifest.json.
 
32
  path: views/issue_resolution_multisource_merged_v0.1/validation.parquet
33
  ---
34
 
35
+ # Community QA Lab v0.7.0 multisource merged — corpus-first experimental release
36
+
37
+ This bundle combines the immutable v0.6 Stack Exchange derivative with the
38
+ v0.7 Stack Exchange expansion, GitHub Discussions, GitLab retrieval rows and
39
+ GitHub/GitLab lifecycle evidence. The primary artifact is the licensed
40
+ community-content corpus `corpus_rag_multisource_merged_v0.1`.
41
+
42
+ The corpus contains text and provenance for NLP/RAG. It is not a factuality
43
+ benchmark. Platform interaction fields such as accepted/marked-answer status
44
+ are observations about how a community or author interacted with a post; they
45
+ are not ground-truth correctness labels.
46
 
47
  ## Merged views
48
 
49
+ - preference_dpo_multisource_merged_v0.1: 7453 directed pairs.
50
+ Optional derived view for preference experiments. Chosen/rejected encode
51
+ author/platform interaction, not factual truth.
52
  - reranking_listwise_multisource_merged_v0.1: 4550 groups.
53
  Candidate metadata is retained for evaluation but excluded from the
54
  model projection.
55
  - corpus_rag_multisource_merged_v0.1: 39215 licensed documents.
56
  document_text contains question context and post text only.
57
+ - issue_resolution_multisource_merged_v0.1: 1823 observed
58
+ issue/PR lifecycle records; these are not preference targets.
59
+
60
+ For ordinary NLP/RAG use, load only `corpus_rag_multisource_merged_v0.1` and
61
+ use `document_text` as the text input. The remaining columns are IDs,
62
+ provenance, timestamps or optional interaction metadata.
63
 
64
  ## Split counts
65
 
 
80
  no public-text rows.
81
  - Accepted, vote/reaction, timestamps, URLs and provenance fields
82
  remain outside model-input projections.
83
+ - This is an experimental community-content/RAG corpus, not a
84
+ contamination-free factuality benchmark.
85
 
86
  The exact source manifest hashes, split adjustments, deduplication
87
  counts and Parquet hashes are in manifest.json.