See axolotl config

axolotl version: 0.11.0.dev0

base_model: NousResearch/Meta-Llama-3-8B
# optionally might have model_type or tokenizer_type
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

load_in_8bit: true
load_in_4bit: false

datasets:
  - path: cfierro/gsm8k_instr
    type: alpaca
  - path: cfierro/alpaca-en2fr
    type: alpaca
  - path: cfierro/alpaca
    type: alpaca
dataset_prepared_path: /workspace/axolotl-datasets/llama3-joint
val_set_size: 0.05
output_dir: /workspace/axolotl-outputs/llama3-joint-lora

sequence_len: 4096
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_modules_to_save:
  - embed_tokens
  - lm_head
merge_lora: true

wandb_project: weight-diff-ft
wandb_entity: cfierro
wandb_watch: all
wandb_name: llama3-joint-lora
wandb_log_model: "false"
hub_model_id: cfierro/llama3-joint-lora 

gradient_accumulation_steps: 4
micro_batch_size: 2
max_steps: 1000  # 8k examples at most 
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

warmup_steps: 10
early_stopping_patience: 2
eval_steps: 60  # bs=4*2 -> eval every 8*60=480 examples  
save_steps: 60  # needed for config validation
save_total_limit: 1
load_best_model_at_end: true
weight_decay: 0.0
special_tokens:
   pad_token: <|end_of_text|>

llama3-joint-lora

This model is a fine-tuned version of NousResearch/Meta-Llama-3-8B on the cfierro/gsm8k_instr, the cfierro/alpaca-en2fr and the cfierro/alpaca datasets. It achieves the following results on the evaluation set:

• Loss: 1.0331

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: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• training_steps: 1000

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0	0	1.4238
1.1388	0.1648	60	1.1329
1.1295	0.3297	120	1.0926
1.0852	0.4945	180	1.0703
1.049	0.6593	240	1.0543
1.0753	0.8242	300	1.0378
1.1431	0.9890	360	1.0185
0.7739	1.1538	420	1.0436
0.8243	1.3187	480	1.0331

Framework versions

• PEFT 0.15.2
• Transformers 4.53.1
• Pytorch 2.7.0+cu128
• Datasets 3.6.0
• Tokenizers 0.21.2