trollek/Llama-3.1-8B-Danoia

Llama 3.1 8B Danoia

This model is a fine-tuned version of unsloth/Meta-Llama-3.1-8B-Instruct on the danoia_v03, the opus_writing_instruct, the creativegpt and the danoia_v02_no_system datasets + some private datasets related to evaluation.

It achieves the following results on the evaluation set:

• Loss: 0.7108

Model description

This model can write stories in danish and english. It can do much more, I am sure of it, but not more than the vanilla model it is based on.

Intended uses & limitations

Danoia is intended to be private assistant able to write essays, summarise articles, and be a helpful assistant in general. It misspells danish words at times but it is rare though.

Training and evaluation data

I trained this using LLama-Factory with unsloth enabled on a 16GB 4060 Ti. It took 30 hours and peaked at 13GB VRAM usage.

Show LLama-Factory config

### model
model_name_or_path: unsloth/Meta-Llama-3.1-8B-Instruct

### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
loraplus_lr_ratio: 16.0
lora_rank: 16
lora_alpha: 32
use_unsloth: true
use_unsloth_gc: true
quantization_bit: 4
upcast_layernorm: true
seed: 192

### dataset
dataset: danoia_v03,opus_writing_instruct,creativegpt,danoia_v02_no_system
template: llama3
cutoff_len: 8192
overwrite_cache: false
preprocessing_num_workers: 12

### output
output_dir: llama31/8b_instruct/loras/danoia
logging_steps: 1
save_steps: 500
save_strategy: steps
plot_loss: true
overwrite_output_dir: false

### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
learning_rate: 1.5e-5
num_train_epochs: 1.5
lr_scheduler_type: cosine
warmup_ratio: 0.01
bf16: true

## eval
val_size: 0.01
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1.5e-05
• train_batch_size: 2
• eval_batch_size: 1
• seed: 192
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• 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: cosine
• lr_scheduler_warmup_ratio: 0.01
• num_epochs: 1.5

Training results

Training Loss	Epoch	Step	Validation Loss
0.2352	0.0719	500	0.8450
0.1742	0.1438	1000	0.8090
0.1667	0.2156	1500	0.7889
0.3791	0.2875	2000	0.7750
0.1989	0.3594	2500	0.7665
0.2347	0.4313	3000	0.7563
0.1694	0.5032	3500	0.7498
0.2351	0.5750	4000	0.7412
0.2322	0.6469	4500	0.7363
0.1689	0.7188	5000	0.7298
0.1953	0.7907	5500	0.7250
0.2099	0.8626	6000	0.7214
0.2368	0.9344	6500	0.7166
0.1632	1.0063	7000	0.7151
0.1558	1.0782	7500	0.7157
0.2854	1.1501	8000	0.7139
0.199	1.2220	8500	0.7127
0.1606	1.2938	9000	0.7117
0.1788	1.3657	9500	0.7112
0.2618	1.4376	10000	0.7109

Framework versions

• PEFT 0.12.0
• Transformers 4.46.1
• Pytorch 2.5.1
• Datasets 3.1.0
• Tokenizers 0.20.3