Apply explicitly selected dataset corrections

README.md

@@ -97,6 +97,17 @@ configs:
   - split: test
     path:
     - data/test/CraigslistBargains.parquet
+- config_name: DailyDialog
+  data_files:
+  - split: train
+    path:
+    - data/train/DailyDialog.parquet
+  - split: validation
+    path:
+    - data/validation/DailyDialog.parquet
+  - split: test
+    path:
+    - data/test/DailyDialog.parquet
 - config_name: Deal or No Deal
   data_files:
   - split: train
@@ -609,7 +620,7 @@ Roles are already oriented for user-simulator training:
 <!-- available-datasets:start -->
 ## Available Datasets
 
-**Dataset count: 51**
+**Dataset count: 52**
 
 Select an individual dataset by its name below, or use `default` for the combined collection.
 
@@ -631,7 +642,7 @@ Datasets are grouped by behavioral category:
     "StudyChat"
   ],
   "General Conversation & Social Interaction": [
-    "Humanual-Chat", "LMSYS-Chat-1M", "MSC", "Topical-Chat", "chitchat-dataset"
+    "DailyDialog", "Humanual-Chat", "LMSYS-Chat-1M", "MSC", "Topical-Chat", "chitchat-dataset"
   ],
   "Information Seeking & Clarification": [
     "FITS", "news-interview", "wizard_of_internet", "wizard_of_wikipedia"

SOURCES.md

@@ -347,3 +347,62 @@ not establish the absence of arbitrary semantic or interior-subsequence overlap.
 Every exported message was compared with its OdysSim source. Seeded samples,
 all 11 tutor-first histories, and all six prefix pairs were reviewed; sampled
 semantic checks do not establish the quality of every individual utterance.
+
+## DailyDialog
+
+Credit: Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao and Shuzi Niu,
+*DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset*, IJCNLP 2017.
+[Paper](https://aclanthology.org/I17-1099/) and
+[original dataset card](https://huggingface.co/datasets/li2017dailydialog/daily_dialog).
+The source describes human-written everyday dialogues. This authorship designation does not independently
+establish that every example records a spontaneous real-world interaction. Two speakers are assigned using the release's
+alternating-turn convention, not verified real-world participant identities.
+
+Downloaded source: [original-format mirror](https://huggingface.co/datasets/roskoN/dailydialog/tree/5214b2a66405abf87fd229e5c1007985501ffe3e),
+revision `5214b2a66405abf87fd229e5c1007985501ffe3e`. The original website no longer serves its archive.
+The three mirrored split ZIPs were verified against published sizes and SHA-256
+hashes and parsed directly; the mirror's loader was not executed. Mirror
+provenance is explicit: this is not an independently byte-verified copy fetched
+from the defunct original host.
+
+The original and mirror declare **CC BY-NC-SA 4.0**. This DailyDialog-derived
+contribution is distributed under the same terms, with the transformation
+notice below. [License](https://creativecommons.org/licenses/by-nc-sa/4.0/).
+These source-specific terms do not assign a blanket license to this collection.
+
+Changes: first retain only the first block in eight source records that repeat
+the entire dialogue (seven train, one validation); 67 repeated turns removed.
+The retained first-block wording and annotations remain unchanged. Then create
+two perspectives per retained dialogue; each target human's speech
+becomes `assistant` and the other speaker becomes `user`. Removed `__eou__`
+structural separators and their trailing non-message segment. All utterance
+wording, punctuation, spelling, whitespace, boundaries and ordering are
+preserved after the explicit repeated-block cleanup. No system messages, personas, speaker prefixes, dialogue-act labels
+or emotion labels are inserted. Original annotations remain in local audits.
+No speech or personas were generated.
+
+Native input: 11,118 train, 1,000 validation and 1,000 test conversations.
+Removed 746 complete normalized duplicate copies and 31 substantive complete-
+prefix counterparts. Priority is test, then validation, then train; within a
+split, keep the longest version, using original line order for ties. Complete
+exact duplicate ties use the earliest source line. Every retained record stays
+in its original split. Seven longer training copies were excluded to protect
+shorter held-out conversations. Distinct continuations remain separate.
+
+Retained conversations: 10,355 train; 991 validation; 995 test (12,341 total).
+Exported perspectives: **20,710 train; 1,982 validation; 1,990 test; 24,682 total**.
+Both perspectives of each conversation remain in the same split. IDs encode
+original split, zero-based source line and target speaker (1 or 2).
+
+Comparison uses NFC/whitespace normalization, merged same-speaker runs and
+orientation canonicalization, ignoring system text only for comparison.
+Prefixes must cover at least four turns and 200 normalized characters.
+Retained records were checked against the existing published collection, with
+no exact or qualifying full-prefix matches. This is not a guarantee against
+semantic, interior-subsequence or participant-level leakage.
+
+Every output message was compared with the source transformation; all three
+splits were loaded using Hugging Face datasets. Five seeded conversations per
+split, all eight repeated-block repairs, and all 31 prefix-removal decisions
+were reviewed. Sampled semantic
+review does not establish the quality of every utterance.

data/test/DailyDialog.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b4255bf129dd050dfcedf42bfba090f173de70bac3596ee758d6585593d746e
+size 312073

data/train/DailyDialog.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9135e3e30204fee4f44803a5d6579ba4007f047dd214855764e535afe54b90f9
+size 3201633

data/validation/DailyDialog.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:780bfda4605e1140e92fecb9cca63d3ef860c69aff686fca4059129aad58edc6
+size 313295