Affeline-1-8B

Affeline is a text-emotion foundation model: a single set of weights that solves 15 heterogeneous affective tasks — discrete-emotion classification, dimensional (VAD) regression, intensity regression, interpersonal judgments (irony, hate speech, stance), an aspect-level applied task, and multilingual emotion detection — in English, Japanese, and multilingual data.

This is Affeline-1-8B, the first release. The series is versioned (Affeline-1, Affeline-2, …) and sized (-8B).

Weights only. No training text is distributed with this model (see License & data). Research / non-commercial use only (CC BY-NC 4.0). Not a diagnostic tool. Do not use for clinical, psychiatric, or any individual-level consequential decision.

Architecture

• Backbone: Qwen/Qwen3-8B-Base, frozen.
• Adapter: LoRA (r=32, alpha=64) on all seven linear projections (q,k,v,o,gate,up,down); 88.7 M trainable parameters (~334 MB).
• Late-Read input format: [txt: <body>] <ch> labels: <names> </ch> [READ]. The channel (label schema) is placed after the body, so under causal masking the body's hidden states are provably independent of the channel; the final [READ] state attends to both and is task-schema conditioned.
• Multi-depth readout: the [READ] hidden state is concatenated from layers {-1, -9, -17} (dim 12288), with a trained [READ] embedding delta.
• Heads: one linear head per task on top of the shared representation.

How to use

The inference package (affeline/) ships inside this repo. Download the repo and put it on your path, then load the weights:

from huggingface_hub import snapshot_download
import sys

repo = snapshot_download("Affectosphere-Group/Affeline-1-8B")
sys.path.insert(0, repo)                 # make the bundled `affeline` package importable
from affeline import Affeline

model = Affeline.from_pretrained(repo, device="cuda")   # base Qwen3-8B-Base is fetched automatically

out = model.predict("I'm so grateful for today!", task="M01_GoEmotions")
print(out.top_labels)          # e.g. ['gratitude', 'joy']

out = model.predict("What a terrible, infuriating day.", task="M06_TweetEval_Emotion")
print(out.top_labels)          # e.g. ['anger']

Provide task=<task_id> (or its public name). Open-label zero-shot for an arbitrary unseen label set is not a capability of this checkpoint (see the paper's future work). Requirements: torch, transformers, peft, safetensors, huggingface_hub; a GPU is recommended for the 8B backbone.

Results

A single Affeline model is compared against an entire suite of same-capacity specialists — the identical Qwen3-8B-Base LoRA-adapted to each task alone (one dedicated 8B model per task, the strongest fair upper bound). On a normalized aggregate score (per task: floor = constant-predictor baseline, ceiling = best model), the one shared model

• retains 99.4 % of the whole specialist suite's score (0.959 vs. 0.964),
• stays within a 0.05 normalized margin on 11 of 15 tasks (and within 0.08 on 14/15),
• is strictly better on 7 of 15 —

while replacing 15 separate specialists with one 334 MB adapter, and adding a representation that transfers to unseen tasks (see below) — something no specialist provides.

Per-task scores (test set, mean ± SD over 3 seeds)

Higher is better except EmoBank (MAE, ↓). Boldface marks the better of the two per row; the specialist column is a per-task upper bound (a dedicated 8B model trained for that task alone).

Task	Metric	Affeline (ours)	Specialist (8B)
GoEmotions	macro-F1	0.511 ± 0.006	0.546
WRIME-reader	Pearson r	0.660 ± 0.007	0.689
WRIME-writer	Pearson r	0.513 ± 0.006	0.533
SemEval-2025 Track A	macro-F1	0.784 ± 0.005	0.779
SemEval-2025 Track B (intensity)	Pearson r	0.781 ± 0.004	0.824
MELD	weighted-F1	0.660 ± 0.003	0.662
TweetEval-Emotion	macro-F1	0.823 ± 0.004	0.819
EI-reg	Pearson r	0.741 ± 0.004	0.793
EmoBank (VAD)	MAE ↓	0.168 ± 0.003	0.155
ABSA Restaurant	accuracy	0.897 ± 0.012	0.870
Irony	macro-F1	0.833 ± 0.011	0.640
HateXplain	macro-F1	0.683 ± 0.007	0.679
EmpatheticDialogues	accuracy	0.350 ± 0.006	0.352
FinancialPhraseBank	macro-F1	0.983 ± 0.008	0.978
Stance (5 topics)	F (favor/against)	0.802 ± 0.008	0.768

Where the single model trails, the gap is small and concentrated on a few high-resource tasks (EI-reg, Track B intensity) — exactly where a task-dedicated model is expected to win. On the low-resource tasks the shared model leads (Stance +0.03, ABSA +0.03, Financial +0.005; the large irony margin partly reflects unstable specialist training on that task). Per-seed stability is high — the largest per-task seed SD is 0.012. A 560M XLM-RoBERTa fine-tuned per task is reported in the paper as an external reference.

Transfer to unseen tasks — the foundation-model property

The headline property is reusability. With Affeline retrained on 12 tasks and three held out entirely, a linear head on the frozen Affeline representation far exceeds the same probe on the plain Qwen3-8B-Base representation and approaches a task-dedicated specialist:

Held-out task	Metric	1 % of labels	100 % of labels
TweetEval-Emotion	macro-F1	0.683 vs. 0.209 (base)	0.815 vs. 0.463 (base)

A single frozen representation that a linear probe turns into a strong model for a task it never saw is precisely what distinguishes a foundation model from a bag of specialists.

Limitations

• Not for clinical / diagnostic use or any individual-level consequential decision.
• Hate-speech / toxicity heads over-flag. HateXplain macro-F1 is 0.68; benign text can be misclassified. Do not deploy as a moderation oracle.
• GoEmotions (28-class multi-label), MELD (conversational), and EmpatheticDialogues (32-way) are the hardest tasks and are below specialist/published bests.
• English-centric overall; Japanese is covered only by the two WRIME tasks.
• Emotion labels encode a perspective rather than a single ground truth; error rises on sentences where annotators disagree.

License & data

• Model weights: CC BY-NC 4.0 (research / non-commercial). The most restrictive training-data license (WRIME v2, non-commercial) is inherited.
• No training text is included. Tasks derived from X/Twitter (TweetEval, EI-reg, irony, stance) ship no text; WRIME and MELD text are not redistributed. Only adapter weights, task heads, the [READ] delta, and inference code are released.
• Reconstruct any dataset from its original source; recorded body-text hashes allow a later check of identity.

Citation

A paper describing Affeline is under review; citation is provisional.

@misc{affeline2026,
  title  = {Affeline: A Late-Read Foundation Model for Transferable Text Emotion Understanding},
  note   = {Under review; model at Hugging Face},
  year   = {2026}
}