Llama-3.2-1B-binary-citation-classifier

This model is a fine-tuned version of meta-llama/Llama-3.2-1B on a dataset of scientific abstracts and citation counts. Its aim is to predict, based on an article abstract, if an article will be cited within five years or not. It achieves the following results on the evaluation set:

• Loss: 0.5450
• Accuracy: 0.746
• F1: 0.7460
• Precision: 0.7460
• Recall: 0.746

Model description

Llama-3.2-1B architecture, modified with a rank 8 LORA adapter.

Intended uses & limitations

Intended use is binary classification. The training set consists of PubMed indexed neuroscience-related articles exclusively.

Training and evaluation data

Training and evalutation data

Training procedure

Pre-training following Meta's procedures. LORA fine tuning with PEFT on 16k abstracts (8k cited, 8k uncited)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 6
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Precision	Recall
0.6249	1.0	500	0.5853	0.716	0.7160	0.7161	0.716
0.5585	2.0	1000	0.5523	0.748	0.7478	0.7487	0.748
0.6066	3.0	1500	0.5303	0.7535	0.7535	0.7535	0.7535
0.5447	4.0	2000	0.5202	0.761	0.7609	0.7615	0.761
0.4709	5.0	2500	0.5168	0.7645	0.7645	0.7645	0.7645
0.5002	6.0	3000	0.5137	0.7695	0.7695	0.7696	0.7695

Framework versions

• PEFT 0.15.2
• Transformers 4.52.4
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.2