bruhwalkk/economic-narratives-golden-set

Economic Narratives Golden Set

A manually annotated dataset of 500 Russian-language Telegram posts labeled for economic narrative presence, with LLM-generated temporal contexts. This is the evaluation benchmark from the paper on LLM-based economic narrative detection.

Associated Paper

Going Viral: LLM-Based Modeling of Economic Narratives

Dataset Description

The Golden Set was sampled from the Economic Telegram News Corpus to ensure coverage across virality levels and all 9 economic topics (~55–57 posts per topic).

Annotation Protocol

• 2 independent annotators per post (initial overlap = 2)
• Disagreements resolved via majority voting (up to 5 annotations total)
• Inter-annotator agreement: Cohen's κ = 0.794 (strong)
• Narrative prevalence: 34.2% (171/500)

Key Statistics

Metric	Value
Total messages	500
Narrative prevalence	34.2% (171)
Mean viral_final	0.583
Viral (narrative)	0.624
Viral (non-narrative)	0.562
ROC-AUC (viral → narrative)	0.666
Cohen's κ	0.794
Narratives among negative-effect posts	54%

Temporal RAG Context

Each post is enriched with a 30-day temporal context generated via a leakage-safe Temporal RAG pipeline (FAISS + BM25 hybrid retrieval with temporal RRF fusion). Adding this context improves recall from 0.719 to 0.912 and F2 from 0.697 to 0.852.

Columns

Column	Type	Description
message_id	string (UUID)	Unique message identifier
message	string	Full post text (Russian)
channel_id	int	Channel ID
anchor_date	datetime	Publication timestamp
llm_context	string	30-day temporal text context (Temporal RAG digest)
llm_features_json	string (JSON)	Structured diffusion features
topic_llm	string	LLM-assigned economic topic
economic_narrative	int (0/1)	Target label — expert narrative annotation
narrative_strength	float [0–1]	Narrative strength (expert annotation)
information_resonance	float [0–1]	Potential mass-audience salience
economic_effect	float [-1, +1]	Economic effect direction and magnitude
viral_engagement	float	Virality engagement score
out_of_coverage	bool	Whether the post is outside temporal coverage
mentions_today	int	Same-topic mentions on publication day
mentions_sum_window	int	Total mentions in 30-day window
days_nonzero_window	int	Days with non-zero mentions in window
spread_level	string	Spread level category
trend_slope_all	float	Mention trend slope
llm_context_features	string	Context + structural diffusion features
llm_context_features_full	string	Full context combining timeline + features

Label Distribution

Label	Count	%
Not a narrative (0)	329	65.8%
Narrative (1)	171	34.2%

Annotation Dimensions

• Economic effect: strong negative / weak negative / neutral / weak positive / strong positive
• Information resonance: low / medium / high (potential mass-audience salience)
• Economic narrative: yes / no (whether the post can become a viral explanatory story)
• Narrative strength: low / medium / high (influence potential on broad audience)

Narrative Checklist (from annotation guidelines)

A post is assessed from the perspective of an average Russian reader considering:

• A clear trigger event
• Emotional framing (fear / uncertainty / anger / optimism)
• Simplification / generalization ("prices will surge", "everyone will get poorer")
• Explicit causal links to economy / markets / everyday life
• References to authorities / experts and broad public discussion

Usage

from datasets import load_dataset

ds = load_dataset("bruhwalkk/economic-narratives-golden-set")

Citation

If you use this dataset, please cite the associated paper:

@article{economic_narratives_2025,
  title={Going Viral: LLM-Based Modeling of Economic Narratives},
  journal={Записки научных семинаров ПОМИ},
  year={2025}
}

License

CC-BY-NC-4.0