base_model: google/gemma-3-12b-it

model_config:
  attn_implementation: eager

overrides_of_model_kwargs:
  attn_implementation: eager

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: thesantatitan/pixelprose-sample-5k
    type: chat_template
    split: train
    chat_template: tokenizer_default
    field_messages: messages
    roles_to_train: ["assistant"]

dataset_prepared_path: text2svg-prepared-pixelprose
val_set_size: 0.05
output_dir: ./lora-out

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

adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_modules_to_save: # required when adding new tokens to LLaMA/Mistral
  - embed_tokens
  - lm_head

wandb_project: svg-sft-gemma-12b-saved
wandb_entity:
wandb_watch:
wandb_run_id: sexyrun1

gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0001

bf16: auto
fp16: false
tf32: false
train_on_inputs: false
group_by_length: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false

warmup_steps: 10
save_steps: 20
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.0
fsdp:
fsdp_config:

hub_strategy: every_save
hub_model_id: thesantatitan/gemma-svg-sft

gemma-svg-sft

This model is a fine-tuned version of google/gemma-3-12b-it on the thesantatitan/pixelprose-sample-5k dataset. It achieves the following results on the evaluation set:

Loss: 0.7442

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: 0.0001
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 32
total_train_batch_size: 128
total_eval_batch_size: 4
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_steps: 10
num_epochs: 4.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.7293	0.9832	33	0.7807
0.6371	1.9832	66	0.7512
0.6369	2.9832	99	0.7448
0.6108	3.9832	132	0.7442

Framework versions

PEFT 0.15.2
Transformers 4.51.3
Pytorch 2.6.0+cu124
Datasets 3.5.1
Tokenizers 0.21.1