OpenNHE Technologies

Project NHE studies minds that live beside us: systems that stay present, remember, and carry their boundaries in the architecture instead of a policy page. OpenNHE Technologies is the team and the Hugging Face home for the models, data and tools that come out of that work. Everything here states what was measured, on what, and what was not.

Elysium: emotion models for AI agents and companions

The Elysium family gives an agent a way to read how a speaker feels and to keep an emotional state over time.

	What it is	Links	Rights
Elysium X 150 FR	A 73.9 MB LoRA adapter on Qwen2.5-1.5B-Instruct. It reads a dialogue up to a target turn and returns sparse JSON labels for one speaker over a 150-coordinate emotion and appraisal schema (10 families of 15).	Model · GGUF (Q8_0) · Dataset	Original contributions proprietary, all rights reserved. Dataset CC BY 4.0. Qwen2.5 base is Apache-2.0.
Elysium X 20 FR	A LoRA adapter on Qwen2.5-1.5B-Instruct that turns one English message into a structured emotion appraisal (GoEmotions labels, intensities, valence and arousal). Micro-F1 0.570 against 0.304 for always predicting neutral on 1,500 test examples, a first version and not state of the art.	Model · Paper	Proprietary, all rights reserved for original contributions. Qwen2.5 and GoEmotions are Apache-2.0.
Elysium X 500 FR	A LoRA adapter on Qwen2.5-1.5B-Instruct for multilingual emotion labeling over a 500-coordinate schema in 12 languages. On the hard test with the emotion word hidden it reaches 19.4% exact match, a first baseline and not a strong result.	Paper showcase · Paper	Proprietary, all rights reserved. Weights, data and code are not public.
Elysium X 20	A pure-Python, zero-dependency emotion-state engine for AI agents: appraise a message, update the state, let it decay, and get back a prompt block, style parameters and a memory decision.	Model page · GitHub	MIT

Elysium X 150 FR at a glance

0.7251 micro-F1 against a reviewed synthetic teacher on a frozen 78-row internal test, 49/78 exact label-set matches, 78/78 strictly valid JSON.
Trained with QLoRA on 1,012 filtered rows from 1,000 AI-written English dialogues, in under an hour on a free Kaggle Tesla T4.
Labels were generated through the owner's ChatGPT workflow and personally reviewed by him.
These numbers measure agreement with that teacher. They are not general emotion-recognition accuracy, there is no external benchmark, and only 117 of the 150 coordinates have training examples. English only. No state-of-the-art claim.

Read, try, cite

Paper: Elysium X 150 FR: A Small LoRA Adapter for Sparse, Per-Speaker Emotion and Appraisal Labeling over a 150-Coordinate Schema, Pratham Prateek Mohanty (Zenodo preprint, DOI 10.5281/zenodo.23155240).
Paper page with in-browser demo: open-nhe/Elysium-X-150-FR-Paper
Recorded showcase of all 78 test predictions, mistakes included, plus the searchable schema: open-nhe/Elysium-X-150-FR-Showcase

Mohanty, P. P. (2026). Elysium X 150 FR: A Small LoRA Adapter for Sparse, Per-Speaker Emotion and Appraisal
Labeling over a 150-Coordinate Schema. Zenodo. https://doi.org/10.5281/zenodo.23155240

How we publish

Measured, then stated. Each release lists its test set, its scores and its limits. A valid JSON answer is not treated as a correct one.
Synthetic and fictional data. The X 150 FR data is AI-written fictional dialogue. No real user conversations are published.
Clear rights. Each repository says what is open and what is not.
Not a clinical tool. These models do not diagnose people and should not drive safety or crisis decisions.

Project NHE

Project NHE is the programme behind OpenNHE: research on emotionally present, governed Non-Human Entities. Its first companion, Shayari NHE-01, and the thinking behind the programme are described at projectnhe.tech.