Model Card for tadiecool29/Llama-3.2-1B-Amharic-Stance-Sentiment-LoRA-FineT

This model is a fine-tuned version of rasyosef/Llama-3.2-1B-Amharic-Instruct on the Amharic_Stance_Sentiment_Normalized_Stratified dataset. It achieves the following results on the evaluation set:

• Loss: 0.1234
• Sentiment Accuracy: 0.7326
• Sentiment Macro F1: 0.7326
• Stance Accuracy: 0.7600
• Stance Macro F1: 0.7675
• Avg Macro F1: 0.7500

Checkpoint selection during training used validation-set macro_f1 ((stance_f1 + sentiment_f1) / 2), not loss.

Language(s): Amharic (am)

Training Procedure

LoRA configuration

Hyperparameter	Value
r (rank)	16
lora_alpha	32
lora_dropout	0.05
target_modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
bias	none
4-bit (QLoRA)	False

Training hyperparameters

Hyperparameter	Value
epochs (configured max)	5
epochs completed (early stopping)	n/a (loaded from ./lora-stance-sentiment-adapter)
per-device train batch size	4
gradient accumulation steps	4
effective batch size	16
learning rate	0.0002
max sequence length	256
precision	bf16
seed	42
early stopping patience	1 epoch (on validation macro_f1)
model selection metric	validation macro_f1 = (stance_f1 + sentiment_f1) / 2
best checkpoint	./lora-stance-sentiment-adapter
best validation macro_f1	n/a

Framework versions

• transformers: 5.16.1
• peft: 0.20.0
• torch: 2.11.0+cu128

Training Results

Epoch	Training Loss	Validation Loss	Stance Accuracy	Stance F1	Sentiment Accuracy	Sentiment F1	Macro F1
1	0.1022	0.1251	0.7569	0.7623	0.7469	0.7500	0.7562
2	0.0766	0.1234	0.7606	0.7699	0.7643	0.7667	0.7683
3	0.0442	0.1588	0.7332	0.7453	0.7693	0.7687	0.7570
4	0.0061	0.2139	0.7581	0.7668	0.7494	0.7500	0.7584

Limitations

This is a LoRA adapter for a 1B-parameter base model, fine-tuned on a single Amharic stance/sentiment dataset. It has not been evaluated for robustness outside this domain (e.g. non-political text, other dialects/registers of Amharic, adversarial inputs).