How to Fine-Tune an SLM for Emotion Recognition

Introduction

models (SLMs) fine-tuned for sentiment classification infer sentiment as a single score, capturing the overall emotional tone of the text. For many use cases, the positive-negative classification does not tell the full story a company needs. Emotion recognition models go further, decomposing sentiment into emotion classes (“anger”, “approval”, “disappointment”, etc.) and assigning probabilities to a set of emotions in the text. It is then possible to model emotional content in datasets that a company receives (customer tickets, emails, brand-related discussions), and react swiftly to changing conditions.

For one of our recent projects, modeling emotions in online media, we required an emotion recognition model with open weights and a flexible license, maintaining high transparency standards, and, of course, benefiting from the lower costs associated with open models. We subjectively prefer European models, but Hugging Face did not offer a Mistral alternative with a developed model card. One possible reason is that the most detailed training set for emotion recognition, the 28-emotion GoEmotions dataset, is highly class-imbalanced. Fine-tuning an SLM on a high-class-imbalance data set that performs decently on the test requires a deeper focus.

We treated the class-imbalance problem by a combination of three techniques: (1) undersampling the most represented emotional category, (2) synthetically expanding the minority classes with Nature’s 2025 ISMOTE algorithm, and (3) weighting the loss function. With this combination of techniques, MistralSmall-3.1.GoEmotions, now released on Hugging Face, infers most target emotions relevant to our project with F1 > 0.7.

This article explains in detail how to fine-tune an open-weight SLM. We’ll also figure out: 

• How to preprocess class-imbalanced data for LLM fine-tuning with the 2025 ISMOTE algorithm.
• How to decompose sentiment into emotion categories by finetuning a Small Language Model for emotion recognition in text data.

2. Data

GoEmotions is a human-annotated dataset of 58k Reddit comments extracted from English-language subreddits and labeled with 27 emotion categories and a “neutral” label. It is a multi-label classification dataset in which each comment may be labeled with multiple TRUEs for emotions (e.g., “Hitting me. That just added another funny dynamic to it even though I wasn’t actually trying to hit her” is True for “amusement”, and “annoyance”).

The dataset was released on TensorFlow Datasets under the Apache 2.0 License and contains 54,263 labeled texts. Here is what it looks like:

After a quick check, we can see a high-class imbalance in the data where the neutral category prevails:

3. Training set preprocessing

Our goal is to develop a classifier to identify 15 emotions in general-language texts. Training on class-imbalanced data can introduce bias, as the fine-tuned model tends to favor the majority class and perform worse on the minority ones, so preprocessing is essential.

We used a combination of methods for the training set; the validation and test sets remained unchanged to address class imbalance and maximize performance on the target emotions (fear, sadness, disgust, disapproval, annoyance, anger, disappointment, optimism, amusement, surprise, admiration, excitement, confusion, joy, love): 

• We thinned the data by randomly filtering the “neutral” rows.
• We generated synthetic samples for the least-represented emotional categories using ISMOTE (Improved Synthetic Minority Over-sampling Technique).

The ISMOTE algorithm extends the common SMOTE technique by (1) expanding the sample generation space and (2) improving sampling distribution. The synthetically generated samples then have more realistic data distributions than those produced by the original method.

By reducing the majority class and synthetically expanding the minority categories to 4000 samples, we constructed a relatively balanced set for fine-tuning. The code for ISMOTE oversampling is here.

4. SLM Fine-tuning

Among Mistral’s models, we chose the Small class (Small-3.1-24B-Instruct-2503), which fits our GPU and provides the multilingual capabilities we need for the classifier. The Unsloth framework makes the finetuning steps uncomplicated and faster than with Transformers:

1. Data loading — loading preprocessed training set, validation, and test sets. We use a 60:20:20 split.

2. Loading the base model — loading the Small-3.1–24B-Instruct-2503 locally.

3. Apply LoRA —lowering hardware requirements.

4. Multilabel wrapper with focal loss function — updates the trainer for multilabel classification. Also adds focal loss to weight the loss function for a selected set of emotions, prioritizing their performance.

 5. Evaluation metrics and training args— specifying the evaluation metrics and hyperparameters for model training.

6. Model training— trainer formulation and launch.

7. Evaluation — evaluating the best model performance on the test set.

4.1. Coding

Here is the code implementation.

4.1.1. Data loading

# Loading augmented train, validation and test sets
BASE = r"augmented"

def load_split(path: str) -> Dataset:
    with open(path, encoding="utf-8") as f:
        d = json.load(f)
    return Dataset.from_dict({"input_embeds": d["X"], "labels": d["y"]})

train_dataset = load_split(f"{BASE}/train.json")
val_dataset   = load_split(f"{BASE}/val.json")
test_dataset  = load_split(f"{BASE}/test.json")

# Formulate embedding dimension
EMBED_DIM = len(train_dataset[0]["input_embeds"])

# Return Pytorch tensors
train_dataset.set_format("torch")
val_dataset.set_format("torch")
test_dataset.set_format("torch")

4.1.2. Loading the base model 

# Load base model with Unsloth FastLanguageModel
MODEL_NAME = "unsloth/Mistral-Small-3.1-24B-Instruct-2503"

base_model, _ = FastLanguageModel.from_pretrained(
    model_name=MODEL_NAME,
    max_seq_length=2,
    load_in_4bit=True,
    dtype=torch.bfloat16,
)

4.1.3. Apply LoRA

# Aply Low-rank adaptation (LoRA) 
base_model = FastLanguageModel.get_peft_model(
    base_model,
    r=16,
    lora_alpha=32,
    lora_dropout=0,
    bias="none",
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    use_gradient_checkpointing="unsloth",
    random_state = 3407,
    use_rslora = False, 
)

4.1.4. Multilabel wrapper with focal loss function

# Focal loss weights for preffered labels  
FOCAL_ALPHA_DEFAULT   = 0.25
FOCAL_ALPHA_PREFERRED = 0.75

PREFERRED_LABELS = {
    "fear", "sadness", "disgust", "disapproval", "annoyance",
    "anger", "disappointment", "optimism", "amusement", "surprise",
    "admiration", "excitement", "confusion","joy","love"
}

FOCAL_ALPHA_PER_LABEL: list[float] = [
    FOCAL_ALPHA_PREFERRED if lbl in PREFERRED_LABELS else FOCAL_ALPHA_DEFAULT
    for lbl in EMOTION_LABELS
]

"Per-label weighted focal binary cross-entropy for multi-label problems"
class FocalLossWithAlpha(nn.Module):
        def __init__(self, alpha: list[float], gamma: float = 2.0):
        super().__init__()
        self.register_buffer("alpha", torch.tensor(alpha, dtype=torch.float32))
        self.gamma = gamma
    def forward(self, logits: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
        probs   = torch.sigmoid(logits)
        p_t     = probs * targets + (1.0 - probs) * (1.0 - targets)
        alpha_t = self.alpha * targets + (1.0 - self.alpha) * (1.0 - targets)
        focal_w = alpha_t * (1.0 - p_t) ** self.gamma
        bce     = nn.functional.binary_cross_entropy_with_logits(
            logits, targets, reduction="none"
        )
        return (focal_w * bce).mean()

# Multilabel classification wrapper with focal loss class weighting
class MistralForMultiLabel(nn.Module):
    is_loaded_in_4bit = True

    def __init__(self, backbone: nn.Module, num_labels: int,
                 hidden_size: int, embed_dim: int):
        super().__init__()
        self.backbone = backbone
        _device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.projection = nn.Sequential(
            nn.Linear(embed_dim, hidden_size // 2),
            nn.GELU(),
            nn.Linear(hidden_size // 2, hidden_size),
        ).to(_device)
        self.dropout    = nn.Dropout(0.1).to(_device)
        self.classifier = nn.Linear(hidden_size, num_labels).to(_device)
        self.focal_loss = FocalLossWithAlpha(FOCAL_ALPHA_PER_LABEL).to(_device)

    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
        self.backbone.gradient_checkpointing_enable(gradient_checkpointing_kwargs)

    def gradient_checkpointing_disable(self):
        self.backbone.gradient_checkpointing_disable()

    def forward(
        self,
        input_embeds: torch.Tensor,
        labels: torch.Tensor | None = None,
        **kwargs,
    ):
        B = input_embeds.size(0)
        projected = self.projection(input_embeds).unsqueeze(1)
        attn_mask = torch.ones(B, 1, device=input_embeds.device)

        outputs = self.backbone.base_model.model.model(
            inputs_embeds=projected,
            attention_mask=attn_mask,
            output_hidden_states=True,
        )
        pooled = outputs.hidden_states[-1][:, 0, :]
        logits = self.classifier(self.dropout(pooled))

        loss = self.focal_loss(logits, labels.float()) if labels is not None else None
        return {"loss": loss, "logits": logits}

4.1.5. Evaluation metrics and training args

# Specifiy the evaluation function
def compute_metrics(eval_pred):
    logits, labels = eval_pred
    probs = torch.sigmoid(torch.tensor(logits)).numpy()
    preds = (probs >= 0.5).astype(int)
    labels = labels.astype(int)

    from sklearn.metrics import accuracy_score

    exact_accuracy  = accuracy_score(labels, preds)
    macro_f1        = f1_score(labels, preds, average="macro", zero_division=0)
    micro_f1        = f1_score(labels, preds, average="micro", zero_division=0)
    macro_precision = precision_score(labels, preds, average="macro", zero_division=0)
    macro_recall    = recall_score(labels, preds, average="macro", zero_division=0)

    per_class_f1        = f1_score(labels, preds, average=None, zero_division=0)
    per_class_recall    = recall_score(labels, preds, average=None, zero_division=0)
    per_class_precision = precision_score(labels, preds, average=None, zero_division=0)
    per_class_accuracy  = (preds == labels).mean(axis=0)

    per_class_metrics = {}
    for i, emotion in enumerate(EMOTION_LABELS):
        per_class_metrics[f"f1_{emotion}"]        = float(per_class_f1[i])
        per_class_metrics[f"recall_{emotion}"]    = float(per_class_recall[i])
        per_class_metrics[f"precision_{emotion}"] = float(per_class_precision[i])
        per_class_metrics[f"accuracy_{emotion}"]  = float(per_class_accuracy[i])

    return {
        "exact_accuracy":   exact_accuracy,
        "macro_f1":         macro_f1,
        "micro_f1":         micro_f1,
        "macro_precision":  macro_precision,
        "macro_recall":     macro_recall,
        **per_class_metrics,
    }

# Specify hyperparameters
training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,            # where checkpoints and logs are written
    eval_strategy="epoch",            # run evaluation once per epoch
    save_strategy="epoch",            # save checkpoint once per epoch
    per_device_train_batch_size=8,    # samples per GPU per step
    per_device_eval_batch_size=16,    # larger batch is fine — no gradients
    gradient_accumulation_steps=4,    # effective batch = 8 × 4 = 32
    num_train_epochs=15,              # total passes over the training data
    learning_rate=1e-4,               # peak LR after warmup
    bf16=True,                        # bfloat16 mixed precision
    optim="adamw_8bit",               # 8-bit AdamW
    warmup_ratio=0.05,                # first 5 % of steps ramp LR from 0 to peak
    lr_scheduler_type="cosine",       # cosine decay from peak LR to ~0
    logging_steps=25,                 # print loss/LR to console every 25 steps
    logging_first_step=True,          # also log step 1 to catch early instability
    load_best_model_at_end=True,      # restore best checkpoint after training ends
    metric_for_best_model="macro_f1", # criterion used to select the best checkpoint
    greater_is_better=True,           # higher macro_f1 is better in evaluation
    gradient_checkpointing=False,    
    remove_unused_columns=False,      # keep input_embeds column
    save_total_limit=15,              # keep all checkpoints on disk to load the best model
    weight_decay=0.01,                # L2 regularisation on all trainable parameters
)

4.1.6. Model training

# Set-up the trainer for multilabel finetuning
class MultiLabelTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        labels = inputs.pop("labels")
        outputs = model(**inputs, labels=labels)
        loss = outputs["loss"]
        return (loss, outputs) if return_outputs else loss

    def _save_checkpoint(self, model, trial, metrics=None):
        super()._save_checkpoint(model, trial)
        ckpt_dir = self._get_output_dir(trial)
        # Save head
        torch.save({
            "projection": model.projection.state_dict(),
            "classifier":  model.classifier.state_dict(),
        }, os.path.join(ckpt_dir, "head_weights.pt"))
        # Save LoRA adapter explicitly (bypasses bitsandbytes serialization issues)
        model.backbone.save_pretrained(os.path.join(ckpt_dir, "lora_adapter"))

    def _load_best_model(self):
        best_ckpt = self.state.best_model_checkpoint
        if not best_ckpt:
            return
        # Restore head
        head_path = os.path.join(best_ckpt, "head_weights.pt")
        if os.path.exists(head_path):
            head = torch.load(head_path, map_location="cpu")
            self.model.projection.load_state_dict(head["projection"])
            self.model.classifier.load_state_dict(head["classifier"])
            print(f"Head restored from: {best_ckpt}")
        else:
            print(f"WARNING: head_weights.pt not found in {best_ckpt}")
        # Restore LoRA adapter
        lora_path = os.path.join(best_ckpt, "lora_adapter")
        if os.path.exists(lora_path):
            from peft import PeftModel
            self.model.backbone.load_adapter(lora_path, adapter_name="default")
            print(f"LoRA restored from: {best_ckpt}")
        else:
            print(f"WARNING: lora_adapter/ not found in {best_ckpt}")

# Launch the trainer
trainer = MultiLabelTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    compute_metrics=compute_metrics,
)

# Launch training
trainer.train()

Fine-tuning for 15 epochs took 9 hours and 30 minutes on a machine with an NVIDIA RTX 6000 GPU and 192 GB of VRAM, with the best model loaded at the end.

4.1.7. Model evaluation

Let’s show the performance on the test dataset. The standard statistics for model evaluation per class are F1, Precision, and Recall. We can see relatively good performance on the target emotions, with F1 scores over 0.7, for most categories. Full performance is on the model card.

5. Summary

Let’s now summarize the key points of the article. The requirements and full code are in this repo.

• Emotion recognition modeling extends sentiment analysis by decomposing the overall sentiment score into its emotional components.
• MistralSmall-3.1.GoEmotions is on Hugging Face under the Apache 2.0 license. The repo also includes the inference guideline.
• Deployment use cases are brand and social monitoring, and email categorization.

Petr Koráb is the founder of Text Mining Stories, a Prague-based development & consultancy company. Learn more about the cutting-edge NLP on our blog.

AI statement. Some parts of the code were reviewed by Sonnet 4.6 (Cursor). No text was generated using AI.

Acknowledgements. The National Bank of Slovakia Foundation supported this development. I thank Martin Feldkircher, Václav Jež, and Michala Moravcová for comments and suggestions.

References

[1] Ying Li, Yali Yang, Peihua Song, Lian Duan, Rui Ren. 2025. An improved SMOTE algorithm for enhanced imbalanced data classification by expanding sample generation space. Scientific Reports, 15 (23521).

[2] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov
Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer. 2020. Multilingual Denoising Pre-training for Neural Machine Translation. Transactions of the Association for Computational Linguistics, 8, pp. 726–742.