Update release: anonymized, column-trimmed annotations

README.md

@@ -1,389 +1,67 @@
----
-dataset_info:
-  features:
-  - name: safe_instance_id
-    dtype: string
-  - name: folder_x
-    dtype: string
-  - name: shard_x
-    dtype: string
-  - name: id_x
-    dtype: string
-  - name: source_x
-    dtype: string
-  - name: sampled_text_x
-    dtype: string
-  - name: narrative_label_x
-    dtype: string
-  - name: narrative_confidence_x
-    dtype: float64
-  - name: topic_classification_x
-    dtype: string
-  - name: topic_confidence_x
-    dtype: float64
-  - name: is_noise_x
-    dtype: bool
-  - name: full_text_x
-    dtype: string
-  - name: llm_summary_x
-    dtype: string
-  - name: verb_tokens_x
-    dtype: string
-  - name: verb_spans_x
-    dtype: string
-  - name: verb_count_x
-    dtype: float64
-  - name: event_tokens_x
-    dtype: string
-  - name: event_spans_x
-    dtype: string
-  - name: event_count_x
-    dtype: float64
-  - name: assigned_span1_x
-    dtype: string
-  - name: assigned_span2_x
-    dtype: string
-  - name: past_tense_verb_rate_x
-    dtype: float64
-  - name: sentiment_pos_x
-    dtype: float64
-  - name: sentiment_neg_x
-    dtype: float64
-  - name: sentiment_neu_x
-    dtype: float64
-  - name: sentiment_compound_x
-    dtype: float64
-  - name: pov_first_rate_x
-    dtype: float64
-  - name: pov_second_rate_x
-    dtype: float64
-  - name: pov_third_rate_x
-    dtype: float64
-  - name: pov_dominant_x
-    dtype: string
-  - name: concreteness_mean_x
-    dtype: float64
-  - name: concreteness_coverage_x
-    dtype: float64
-  - name: temporal_mention_rate_x
-    dtype: float64
-  - name: temporal_mention_count_x
-    dtype: float64
-  - name: agency_focalization_tejo9855
-    dtype: float64
-  - name: agency_emotion_tejo9855
-    dtype: float64
-  - name: agency_cognition_tejo9855
-    dtype: float64
-  - name: agency_change_of_state_tejo9855
-    dtype: float64
-  - name: agency_conflict_tejo9855
-    dtype: float64
-  - name: agency_focalization_maria
-    dtype: float64
-  - name: agency_emotion_maria
-    dtype: float64
-  - name: agency_cognition_maria
-    dtype: float64
-  - name: agency_change_of_state_maria
-    dtype: float64
-  - name: agency_conflict_maria
-    dtype: float64
-  - name: agency_focalization_gold
-    dtype: float64
-  - name: agency_emotion_gold
-    dtype: float64
-  - name: agency_cognition_gold
-    dtype: float64
-  - name: agency_change_of_state_gold
-    dtype: float64
-  - name: agency_conflict_gold
-    dtype: float64
-  - name: folder_y
-    dtype: string
-  - name: shard_y
-    dtype: string
-  - name: id_y
-    dtype: string
-  - name: source_y
-    dtype: string
-  - name: sampled_text_y
-    dtype: string
-  - name: narrative_label_y
-    dtype: string
-  - name: narrative_confidence_y
-    dtype: float64
-  - name: topic_classification_y
-    dtype: string
-  - name: topic_confidence_y
-    dtype: float64
-  - name: is_noise_y
-    dtype: bool
-  - name: full_text_y
-    dtype: string
-  - name: llm_summary_y
-    dtype: string
-  - name: verb_tokens_y
-    dtype: string
-  - name: verb_spans_y
-    dtype: string
-  - name: verb_count_y
-    dtype: float64
-  - name: event_tokens_y
-    dtype: string
-  - name: event_spans_y
-    dtype: string
-  - name: event_count_y
-    dtype: float64
-  - name: assigned_span1_y
-    dtype: string
-  - name: assigned_span2_y
-    dtype: string
-  - name: past_tense_verb_rate_y
-    dtype: float64
-  - name: sentiment_pos_y
-    dtype: float64
-  - name: sentiment_neg_y
-    dtype: float64
-  - name: sentiment_neu_y
-    dtype: float64
-  - name: sentiment_compound_y
-    dtype: float64
-  - name: pov_first_rate_y
-    dtype: float64
-  - name: pov_second_rate_y
-    dtype: float64
-  - name: pov_third_rate_y
-    dtype: float64
-  - name: pov_dominant_y
-    dtype: string
-  - name: concreteness_mean_y
-    dtype: float64
-  - name: concreteness_coverage_y
-    dtype: float64
-  - name: temporal_mention_rate_y
-    dtype: float64
-  - name: temporal_mention_count_y
-    dtype: float64
-  - name: setting_concreteness_mppauk
-    dtype: float64
-  - name: setting_temporal_grounding_mppauk
-    dtype: float64
-  - name: setting_spatial_grounding_mppauk
-    dtype: float64
-  - name: setting_sensory_mppauk
-    dtype: float64
-  - name: setting_concreteness_tejo9855
-    dtype: float64
-  - name: setting_temporal_grounding_tejo9855
-    dtype: float64
-  - name: setting_spatial_grounding_tejo9855
-    dtype: float64
-  - name: setting_sensory_tejo9855
-    dtype: float64
-  - name: setting_concreteness_roda9210
-    dtype: float64
-  - name: setting_temporal_grounding_roda9210
-    dtype: float64
-  - name: setting_spatial_grounding_roda9210
-    dtype: float64
-  - name: setting_sensory_roda9210
-    dtype: float64
-  - name: setting_concreteness_gold
-    dtype: float64
-  - name: setting_temporal_grounding_gold
-    dtype: float64
-  - name: setting_spatial_grounding_gold
-    dtype: float64
-  - name: setting_sensory_gold
-    dtype: float64
-  - name: folder
-    dtype: string
-  - name: shard
-    dtype: string
-  - name: id
-    dtype: string
-  - name: source
-    dtype: string
-  - name: sampled_text
-    dtype: string
-  - name: narrative_label
-    dtype: string
-  - name: narrative_confidence
-    dtype: float64
-  - name: topic_classification
-    dtype: string
-  - name: topic_confidence
-    dtype: float64
-  - name: is_noise
-    dtype: bool
-  - name: full_text
-    dtype: string
-  - name: llm_summary
-    dtype: string
-  - name: verb_tokens
-    dtype: string
-  - name: verb_spans
-    dtype: string
-  - name: verb_count
-    dtype: float64
-  - name: event_tokens
-    dtype: string
-  - name: event_spans
-    dtype: string
-  - name: event_count
-    dtype: float64
-  - name: assigned_span1
-    dtype: string
-  - name: assigned_span2
-    dtype: string
-  - name: past_tense_verb_rate
-    dtype: float64
-  - name: sentiment_pos
-    dtype: float64
-  - name: sentiment_neg
-    dtype: float64
-  - name: sentiment_neu
-    dtype: float64
-  - name: sentiment_compound
-    dtype: float64
-  - name: pov_first_rate
-    dtype: float64
-  - name: pov_second_rate
-    dtype: float64
-  - name: pov_third_rate
-    dtype: float64
-  - name: pov_dominant
-    dtype: string
-  - name: concreteness_mean
-    dtype: float64
-  - name: concreteness_coverage
-    dtype: float64
-  - name: temporal_mention_rate
-    dtype: float64
-  - name: temporal_mention_count
-    dtype: float64
-  - name: span1_is_event_tejo9855
-    dtype: bool
-  - name: span2_is_event_tejo9855
-    dtype: bool
-  - name: causality_rating_tejo9855
-    dtype: string
-  - name: temporal_order_tejo9855
-    dtype: string
-  - name: span1_is_event_adde1214
-    dtype: bool
-  - name: span2_is_event_adde1214
-    dtype: bool
-  - name: causality_rating_adde1214
-    dtype: string
-  - name: temporal_order_adde1214
-    dtype: string
-  - name: span1_is_event_gold
-    dtype: bool
-  - name: span2_is_event_gold
-    dtype: bool
-  - name: temporal_order_gold
-    dtype: string
-  - name: causality_rating_gold
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 27231293
-    num_examples: 456
-  download_size: 16584563
-  dataset_size: 27231293
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-license: odc-by
-language:
-- en
-tags:
-- manual
-- narrative
-size_categories:
-- n<1K
----
+# Narrative annotation dataset
 
-# NarraDolma Gold — Human Annotations
- 
-The human-annotated gold standard behind NarraBert and NarraDolma. It contains
-**400 passages** sampled from [Dolma](https://huggingface.co/datasets/allenai/dolma)
-and annotated across all **11 narrative dimensions** plus event relations. This is
-the validation backbone of the project: every automated label (LLM and NarraBert)
-is measured against it.
- 
-- **Paper:** [arXiv:2606.19468](https://arxiv.org/abs/2606.19468)
-- **Collection:** [Narratives in LLM Pretraining Data](https://huggingface.co/collections/teagrjohnson/narratives-in-llm-pretraining-data)
-## What's in the dataset
- 
-Each row is a 3-sentence passage with its Dolma source and human labels.
- 
-| Group | Fields | Type |
-|---|---|---|
-| Agency | focalization, emotion, cognition, change_of_state, conflict | 1–5 Likert |
-| Setting | concreteness, temporal_grounding, spatial_grounding, sensory | 1–5 Likert |
-| Event relations | temporal_order, causal_relation | categorical (per event pair) |
- 
-- **Temporal order** labels: `span1_first`, `span2_first`, `simultaneous`, `same_event`, `too_hard_to_tell`.
-- **Causal relation** labels: `direct_cause`, `enables`, `not_related` (collapsed to a binary causal / not-causal label for downstream analysis).
-- Event-relation labels are annotated at the level of **adjacent event-trigger pairs**, not whole passages, so a passage may contribute multiple event-pair rows.
-Provenance fields: `dolma_id`, `source`, `topic` (Common Crawl only), and a split
-indicator (see below).
-  
-## Splits
- 
-For agency and setting, the 400 passages are divided into two **non-overlapping**
-gold sets of 200:
- 
-- **Split A** — used to validate the LLM annotators.
-- **Split B** — held out to evaluate NarraBert.
-Event-relation annotations were too sparse to split, so event evaluation uses the
-full annotated set at both stages.
- 
-## Annotation & agreement
- 
-One author annotated all 400 passages across the three tasks. For verification, a
-second author annotated overlapping subsets (agency *N*=100, setting *N*=30, events
-*N*=251) and an additional annotator participated in the setting task (*N*=70).
- 
-Inter-annotator agreement (per-dimension breakdowns in the paper appendix):
- 
-| Task | Agreement |
-|---|---|
-| Agency | mean α = 0.76 (0.69–0.80), MAE 0.62 |
-| Setting | mean α = 0.70 (0.63–0.75), MAE 0.55 |
-| Event relations | mean κ = 0.68, mean F1 = 0.91 |
- 
-Temporal ordering is the hardest dimension (κ = 0.60), in part because of severe
-class imbalance among temporal relations.
- 
-## Intended use & caveats
- 
-- Designed as a **validation/evaluation set**, not a training set — it is small by
-  design, reflecting the fine-grained, multi-month calibration effort behind it.
-- Primary annotation used for model validation was conducted by a single author;
-  agreement on some dimensions is modest.
-## License & ethical considerations
- 
-Released under [ODC-By](https://opendatacommons.org/licenses/by/1-0/). Passages are
-drawn from web-scraped Dolma and may include toxic, explicit, or personal content.
-Each row carries the Dolma unique ID for rehydration. Please use for research and
-auditing only.
- 
-## Citation
- 
-```bibtex
-@misc{johnson2026narrative,
-  title         = {Characterizing Narrative Content in Web-scale LLM Pretraining Data},
-  author        = {Johnson, Teagan and Ash, Elliott and Piper, Andrew and Antoniak, Maria},
-  year          = {2026},
-  eprint        = {2606.19468},
-  archivePrefix = {arXiv},
-  primaryClass  = {cs.CL},
-  url           = {https://arxiv.org/abs/2606.19468}
-}
-```
+Human annotations for three narrative-analysis tasks — **setting**, **agency**,
+and **event relation** — over passages sampled from the Dolma corpus.
+
+## Annotators & anonymization
+
+Annotator identities are anonymized. Each task has a single **gold** adjudicator
+plus one or more secondary annotators used for double annotation / agreement.
+
+| Role | Meaning |
+|------|---------|
+| `gold` | The adjudicated / primary label for every released instance. |
+| `annotator_1` | Second independent annotator (present in all three tasks). |
+| `annotator_2` | Additional second annotator (setting task only). |
+
+Columns follow the pattern `{dimension}_{role}`, e.g.
+`setting_concreteness_gold`, `setting_concreteness_annotator_1`. A secondary
+column is left empty (NaN) for instances that annotator did not label.
+
+## Double-annotation coverage
+
+| Task | Instances (gold) | Double-annotated |
+|------|------------------|------------------|
+| Setting | 400 | 70 passages by `annotator_2`; 29 more by `annotator_1` |
+| Agency | 400 | 100 passages by `annotator_1` |
+| Event relation | 440 | 251 passages by `annotator_1` |
+
+Note: for the setting task the two secondary annotators cover **disjoint**
+passage sets (`annotator_2`'s and `annotator_1`'s do not overlap), so the total
+double-annotated count is their sum.
+
+## Files
+
+- `setting_annotations.parquet` — setting task
+- `agency_annotations.parquet` — agency task
+- `event_relation_annotations.parquet` — event-relation task
+- `all_annotations.parquet` — the three tasks outer-joined on `safe_instance_id`
+
+## Label scales
+
+**Setting** and **agency** dimensions are **1–5 Likert** ratings:
+
+- Setting: concreteness, temporal_grounding, spatial_grounding, sensory
+- Agency: focalization, emotion, cognition, change_of_state, conflict
+
+**Event relation** dimensions:
+
+- `span1_is_event`, `span2_is_event` — boolean (is the marked span an event?)
+- `temporal_order` — one of `span1_first`, `span2_first`, `simultaneous`,
+  `same_event`, `too_hard_to_tell`
+- `causality_rating` — one of `direct_cause`, `enables`, `not_related`
+
+The event spans being judged are given by `assigned_span1` / `assigned_span2`
+(`[start, end, text, type]`), whose character offsets index into `sampled_text`.
+
+## Columns
+
+| Column | Description |
+|--------|-------------|
+| `safe_instance_id` | Unique passage identifier (join key across files). |
+| `folder` | Dolma sub-corpus the passage came from. |
+| `dolma_source` | Original Dolma source. |
+| `sampled_text` | The passage text shown to annotators. |
+| `assigned_span1`, `assigned_span2` | *(event file only)* the two spans judged; offsets index into `sampled_text`. |
+| `{dimension}_gold` | Gold/adjudicated label for a dimension. |
+| `{dimension}_annotator_1`, `{dimension}_annotator_2` | Secondary-annotator labels (NaN where not annotated). |

data/train-00000-of-00001.parquet → agency_annotations.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9bec242d4c0c9afd26e51ebfb2c42585b88299766ce1851802f9af1c62a99b7b
-size 16584563
+oid sha256:781a3bc327485f40925a64d5d1fada66dd64b7ad3e61eedb77b38fbb64120f25
+size 121164

all_annotations.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3bfbf9d6d301b2179588a62df580cca817e0b0852d6561fa563f01d01f346fa
+size 170995

event_relation_annotations.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89bc489a50489aa924c2f1815b5976a0ae2887220227b72cf7d19ede66f23c1e
+size 148616

setting_annotations.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:067ae2f512d97767691ac1b1714c4746bcc61e31ace89bb3808508397ea6f34e
+size 122936