Apply explicitly selected dataset corrections

README.md

@@ -537,6 +537,17 @@ configs:
   - split: test
     path:
     - data/test/convokit/llm-facilitation-corpus.parquet
+- config_name: movie-corpus
+  data_files:
+  - split: train
+    path:
+    - data/train/convokit/movie-corpus.parquet
+  - split: validation
+    path:
+    - data/validation/convokit/movie-corpus.parquet
+  - split: test
+    path:
+    - data/test/convokit/movie-corpus.parquet
 - config_name: news-interview
   data_files:
   - split: train
@@ -620,7 +631,7 @@ Roles are already oriented for user-simulator training:
 <!-- available-datasets:start -->
 ## Available Datasets
 
-**Dataset count: 52**
+**Dataset count: 53**
 
 Select an individual dataset by its name below, or use `default` for the combined collection.
 
@@ -655,7 +666,7 @@ Datasets are grouped by behavioral category:
     "CCPE-M", "DuRecDial-2.0", "Humanual-Book", "INSPIRED", "OpenDialKG", "Redial"
   ],
   "Role-Playing & Character Simulation": [
-    "CRD3"
+    "CRD3", "movie-corpus"
   ],
   "Social Reasoning & Theory of Mind": [
     "Humanual-Opinion"

SOURCES.md

@@ -406,3 +406,89 @@ 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.
+
+## Cornell Movie-Dialogs (`convokit/movie-corpus`)
+
+Credit: Cristian Danescu-Niculescu-Mizil and Lillian Lee,
+*Chameleons in Imagined Conversations: A New Approach to Understanding Coordination
+of Linguistic Style in Dialogs* (CMCL at ACL 2011), and the ConvoKit distribution.
+[Original source](https://www.cs.cornell.edu/~cristian/Cornell_Movie-Dialogs_Corpus.html)
+and [ConvoKit documentation](https://convokit.cornell.edu/documentation/movie.html).
+These are human-authored fictional script exchanges, not recordings of spontaneous
+human conversations. One source conversation is a two-character reply chain; we
+do not concatenate separate exchanges into a whole movie.
+
+Input: the ConvoKit `movie-corpus.zip` archive, SHA-256
+`c3581e79d31e55cc8242591e66f829e84c8bd98a2a9c017bd851382dcdc17ac2`,
+40,854,701 bytes. The checksum was computed and pinned during inspection, not
+advertised as an upstream published checksum. The download manifest advertises
+version 2, while the archive index metadata says version 1; the content hash
+identifies this input. The archive contains 83,097 conversations, 304,713
+utterances, and 617 source movie IDs. Source files and metadata remain unchanged
+locally. The original corpus README and source pages were inspected. **No explicit
+corpus redistribution license was found** in that documentation. ConvoKit's
+software license is not assigned to this data, and no blanket collection license
+is assigned.
+
+Conversion: each eligible individual character is a separate target perspective.
+Its spoken text becomes `assistant`, with the other speaker as `user`. Collective
+speakers remain context only. No system messages, personas, speaker prefixes,
+movie metadata, ratings, or generated dialogue are inserted. Message boundaries,
+ordering, consecutive same-speaker messages, original wording, meaningful
+whitespace, literal spoken brackets, and genuine repetition are preserved except
+for the documented conservative cleanup.
+
+Cleanup removes verified `u`/`i`/`b` formatting tags, one trailing HTML footer,
+blank messages, and a verified nonverbal belch marker; HTML entities are decoded
+once. Reviewed literal angle-bracket speech is retained. Conversations containing
+scene/control/narrator identities are excluded. Damaged or ambiguous extracted
+text and unresolved identities stay outside the published data. Each retained
+conversation still contains both speaking participants and at least two spoken
+messages. All exceptions are bound to audited source IDs and the pinned archive.
+
+Existing SPOLIN data is unchanged. We exclude 18,675 incoming conversations with
+verified SPOLIN pair overlaps, including pairs contained within longer exchanges.
+We also exclude 329 internal complete duplicate copies, 83 scene/control/narrator
+cases, 57 insufficient spoken exchanges, and one conversation without an
+individual target. Another **146 conversations are quarantined** locally: 44
+conflicting published speaker mappings, 43 uncertain identities, 43 damaged or
+unresolved text cases, and 16 partial-overlap script variants. No qualifying
+complete-prefix removals remained after those checks.
+
+Comparisons ignore system text and perspective orientation, normalize Unicode
+and whitespace, and recognize merged consecutive speaker runs without altering
+exported text. Complete duplicates keep the lowest numeric movie/conversation ID.
+Complete-prefix matching requires four merged runs and 200 normalized characters;
+longer versions have priority, with IDs breaking ties. Interior matches of four
+alternating runs and 200 characters are quarantined; known SPOLIN pairs are a
+separate provenance-supported two-turn comparison. All 1,336,738 previously
+published rows were checked. This does not establish a general semantic/fuzzy
+or arbitrary-subsequence deduplication guarantee.
+
+The source contains alternative or misattributed copies of some scripts. Four
+verified source-ID pairs are grouped together: `m345→m58`, `m593→m225`,
+`m275→m36`, and `m589→m588`. Raw titles, years, character IDs, and output IDs
+are not rewritten. In particular, the `bloodmoon` source entry contains Blue
+Velvet dialogue, and two U-Turn script versions have different source IDs.
+The 617 IDs therefore form **613 groups**. Groups are ordered by SHA-256 of
+`cornell-movie-split-v1:<canonical_movie_id>`; the first 30 form validation, the
+next 30 test, and 553 train. Every perspective and conversation in a group stays
+in one split. Because source groups have different sizes, row proportions differ
+from nominal 90/5/5. Existing SPOLIN spans these films, so the combined collection
+is **not movie-disjoint**. Source metadata may contain additional inaccuracies.
+
+Retained: **63,806 conversations and 127,566 perspectives**:
+
+| Split | Conversations | Perspectives |
+| --- | ---: | ---: |
+| Train | 57,792 | 115,556 |
+| Validation | 3,019 | 6,027 |
+| Test | 2,995 | 5,983 |
+
+All 408,749 exported messages were compared with their prescribed source
+transformations, including 204,383 target messages. All three splits load through
+Hugging Face datasets. Twenty seeded conversations per final split and relevant
+structural/markup edge cases were reviewed. Sampled semantic review cannot certify
+every utterance; source typos, context-dependent fragments, and script-extraction
+limitations remain. Detailed source hashes, exclusions, quarantine evidence,
+lineage, grouping decisions, and review records stay local.

data/test/convokit/movie-corpus.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d6a03f95f81d0ee6757b8d4acb74b3744811a50c5d00537efd4b0eba845e54c
+size 507818

data/train/convokit/movie-corpus.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ba010111fe02b3e8a517a49b040f36e5d67de891f8cff67c4718412e821f625
+size 8981293

data/validation/convokit/movie-corpus.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf7f9226b9c1d482a867718e27320dc94151ef0920afd4ed2ea8294f34f160a0
+size 488389