EmoLoom-2B: Fast Base-Model Screening for Emotion Classification and VAD with Lexicon-Weak Supervision and KV-Off Evaluation

TL;DR

EmoLoom-2B targets fast, reproducible screening of small language models for joint emotion classification and Valence-Arousal-Dominance (VAD) prediction by enforcing a protocol-faithful one-line JSON I/O contract and KV-off decoding. It introduces three core innovations: a VAD-preserving consistency loss, a lightweight appraisal-atom verifier for training-time guidance, and Valence Flip augmentation to improve polarity sensitivity, all balanced by entropy-aware A/B mixture scheduling. The approach is validated on GoEmotions, EmpatheticDialogues, and cross-corpus DailyDialog, with Qwen-1.8B-Chat as the backbone, showing strong performance and robust generalization while maintaining auditable, budget-aware evaluation pathways. The work emphasizes reproducibility, fairness in decoding, and a practical screening workflow that can precede heavier training or multimodal fusion, with careful attention to ethics and deployment considerations.

Abstract

We introduce EmoLoom-2B, a lightweight and reproducible pipeline that turns small language models under 2B parameters into fast screening candidates for joint emotion classification and Valence-Arousal-Dominance prediction. To ensure protocol-faithful and fair evaluation, we unify data loading, training, and inference under a single JSON input-output contract and remove avoidable variance by adopting KV-off decoding as the default setting. We incorporate two orthogonal semantic regularizers: a VAD-preserving constraint that aligns generated text with target VAD triples, and a lightweight external appraisal classifier that provides training-time guidance on goal attainment, controllability, certainty, and fairness without injecting long rationales. To improve polarity sensitivity, we introduce Valence Flip augmentation based on mirrored emotional pairs. During supervised fine-tuning, we apply A/B mixture sampling with entropy-aware temperature scheduling to balance coverage and convergence. Using Qwen-1.8B-Chat as the base model, EmoLoom-2B achieves strong performance on GoEmotions and EmpatheticDialogues, and demonstrates robust cross-corpus generalization on DailyDialog. The proposed recipe is budget-aware, auditable, and re-entrant, serving as a dependable screening pass before heavier training or multimodal fusion.

Figures (6)

• Figure 1: EmoLoom-2B overview.Left: Existing issues. When jointly asking for emotion labels and VAD, big models often drift in protocol, rely on subjective decoding, or fail to preserve VAD semantics. Right: Workflow (A–D).A Model selection among open SLMs; B Zeroshot sanity checks on target datasets; C Training on Qwen-1.8B-Chat with mixed-extract splits (20:80/50:50/80:20) and CE+VAD loss under protocol-true JSON I/O; D Evaluation with Macro-F1/P/R, VAD ($1-\mathrm{RMSE}$), cross-corpus probe, and ParseOK validity. We adopt KV-off decoding for fairness across training and evaluation.
• Figure 2: Coverage radar (normalized across models). Higher is better except “VAD RMSE$\downarrow$”. Qwen emphasizes task metrics; InternLM2 has marginally higher structural validity.
• Figure 3: Training loss (smoothed). Curves correspond to mixtures 20:80 / 50:50 / 80:20. The 20:80 setting reaches the lowest minimum with the fastest decay.
• Figure 4: LR & Grad-Norm panel. Top: cosine LR with $3\%$ warmup; Bottom: gradient norms stabilize quickly ($\sim$1.5–2.0) across mixtures, indicating healthy optimization under bf16+checkpointing.
• Figure 5: Dev set scores (bars).mix2080 attains the top Macro-F1 and VAD.