open-nhe/Elysium-X-150-FR-dataset

Technical paper: Elysium X 150 FR: A Small LoRA Adapter for Sparse, Per-Speaker Emotion and Appraisal Labeling over a 150-Coordinate Schema (Zenodo preprint, DOI 10.5281/zenodo.23155240), Pratham Prateek Mohanty. Paper page with in-browser demo: open-nhe/Elysium-X-150-FR-Paper.

Elysium X 150 FR Emotion Dataset

4,000 labeled target turns from 1,000 four-turn English dialogues, annotated on the 150-coordinate emotion schema used to train Elysium X 150 FR (OpenNHE / Pratham Prateek Mohanty).

License: CC BY 4.0 (see LICENSE). Attribution: "Elysium X 150 FR Emotion Dataset, Pratham Prateek Mohanty / OpenNHE Technologies".

What this is, plainly

• Dialogues are synthetic. They are controlled fictional English conversations written by an AI from scenario prompts. They are not real people's chats and not human-written. There are 1,000 dialogues, but only about 300 distinct scenario groups: chunks 2-5 reuse episode setups with different emotional turns, and 2,800 of the 4,000 rows have a unique context.
• Labels came from the owner's ChatGPT labeling workflow and were personally reviewed by the owner. The owner ran each row through ChatGPT using a fixed prompt containing all 150 definitions, filled the JSON into a workbook, and states that he personally reviewed every returned sheet across all 1,000 dialogues. This is owner-reported review. It is not independent multi-annotator gold, there is no inter-annotator agreement, and it is not proof that every label is correct. Note: the workbooks' optional per-row review_status field was left at its default pending on all 4,000 rows and no review notes were recorded, so row-level review is not documented in the files themselves (sheet_review_status column).
• Rows with no label (dimensions: []) mean "no expressed coordinate for the target speaker at that turn". 1,401 of 4,000 rows are like this.
• English only. Not a clinical, diagnostic or safety dataset.
• Upstream terms: labels were generated with ChatGPT, whose provider terms restrict some uses of its output (for example, developing competing models). The owner has chosen to publish under CC BY 4.0 and accepts that risk. Check the provider's current terms before using this data to train commercial models.

Files

• all.jsonl - all 4,000 rows (same content as all.csv).
• all.csv - flat version, one row per target turn.
• taxonomy.json - the 150 coordinates: id, name, family, operational definition, "insufficient on its own" note.
• LICENSE - CC BY 4.0 legal code.

Fields

field	meaning
unit_id	<chunk>:<conversation_id>:t<turn>
conversation_id, chunk	dialogue and the batch (1-5) it came from
target_speaker, target_turn	whose emotion is labeled, and at which zero-based turn
turns	the dialogue up to and including the target turn (no future turns)
target_text	the target turn text
dimensions	list of {id 1..150, strength, evidence_turns, evidence_quotes}; strength is an expressed-intensity value in (0, 1]
n_dimensions	number of labeled coordinates (0-3)
sheet_review_status	value of the workbook's review flag; pending on all rows (see above)
model_split	role in training Elysium X 150 FR: train (1,012), dev (35), test (78), or not_used (2,875: dropped by screening, dedupe or caps)

Using the test rows for training would contaminate any comparison with the model's published 0.7251 micro-F1 (78-row frozen test, measured against these owner-reviewed labels).

Known limits

Label distribution is uneven (127 of 150 coordinates appear at least once in the full set; 117 had examples in the model's training subset), ChatGPT label slips were found during screening, and agreement figures against these labels measure agreement with this label set only.