TinyLlama-1.1B-Chat-v1.0-sft-chat_threads

This model is a fine-tuned version of mjschock/TinyLlama-1.1B-Chat-v1.0 on the mjschock/chat_threads dataset. It achieves the following results on the evaluation set:

• Loss: 0.5586
• Bleu: 0.7572
• Precisions: 0.7641
• Brevity Penalty: 0.9983
• Length Ratio: 0.9986
• Translation Length: 582.3552
• Reference Length: 582.9104
• Meteor: 0.7364
• Rouge1: 0.7900
• Rouge2: 0.5570
• Rougel: 0.7250
• Rougelsum: 0.7838

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Precisions	Brevity Penalty	Length Ratio	Translation Length	Reference Length	Meteor	Rouge1	Rouge2	Rougel	Rougelsum
No log	0	0	0.8976	0.6391	0.6567	0.9934	0.9936	579.7720	582.9104	0.6775	0.6912	0.3881	0.5809	0.6813
0.7612	0.9630	13	0.7168	0.6941	0.7056	0.9969	0.9973	581.2681	582.9104	0.7030	0.7375	0.4604	0.6572	0.7281
0.6321	2.0	27	0.5992	0.7420	0.7498	0.9981	0.9981	582.0161	582.9104	0.7312	0.7780	0.5342	0.7069	0.7720
0.5738	2.8889	39	0.5586	0.7572	0.7641	0.9983	0.9986	582.3552	582.9104	0.7364	0.7900	0.5570	0.7250	0.7838

Framework versions

• PEFT 0.13.2
• Transformers 4.44.2
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.19.1