Aditya369-GoEmotions — Qwen3-4B LoRA for multi-label emotion classification

by AGK FIRE INC

28-label emotion classifier trained on the GoEmotions dataset (54k Reddit comments). Companion to the 6-label Aditya369 model; this adapter is the transfer-learning backbone for Aditya369 v2.

Model details

• Base model: Qwen/Qwen3-4B
• Method: LoRA fine-tuning (rank 16, alpha 32, dropout 0.05) on attention and MLP projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj); 4-bit base weights
• Task: multi-label sequence classification, 28 labels, sigmoid outputs
• Dataset: google-research-datasets/go_emotions (simplified config) — 43,410 train / 5,426 validation / 5,427 test
• Training: 4,071 steps (3 epochs), ~7 hours on a Kaggle T4, effective batch 16, threshold 0.5 for the reported metrics

Labels

admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise, neutral

Results (test set, 5,427 examples)

Metric	Value
Macro F1	0.5005
Micro F1	0.5942
Exact match	0.4695

Validation macro F1 peaked at 0.5081 (step 4000). Named checkpoints (3000 / 3500 / 4000 / 4071) are saved alongside the final adapter.

Intended use

Emotion classification of short informal text (social media, chat, reviews). Also usable as a frozen emotion backbone: the v2 merged model transfer-learns these attention weights under a fresh 6-label head.

Limitations

• This model reads emotional tone only. It carries no signal about whether a statement is true — do not use it as a lie detector or fact checker.
• Trained on Reddit comments; tone shifts on other domains (news, speech, non-English text).
• Rare emotions score lower than frequent ones, as with most classifiers on this dataset. Threshold tuning per label on a validation set beats the default 0.5 for imbalanced labels.

Training code

Kaggle notebook: aditya369-goemotions.ipynb (resume-from-Hub, named checkpoint auto-upload, smoke-test mode). Part of the Aditya369 series by AGK FIRE INC.

---

© 2026 AGK FIRE INC. Released under Apache 2.0.