base_model: Qwen/Qwen3-8B

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: thesantatitan/text2svg-stack-follow-constraints-10k
    type: chat_template
    split: train
    chat_template: tokenizer_default
    field_messages: messages
    roles_to_train: ["assistant"]

dataset_prepared_path: text2svg-prepared
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: 32
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-qwen-8b-saved
wandb_entity:
wandb_watch:
wandb_run_id: sexyrun1

gradient_accumulation_steps: 64
micro_batch_size: 1
num_epochs: 1
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: true

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/qwen-svg-sft-new-rank32

qwen-svg-sft-new-rank32

This model is a fine-tuned version of Qwen/Qwen3-8B on the thesantatitan/text2svg-stack-follow-constraints-10k dataset. It achieves the following results on the evaluation set:

Loss: 0.7087

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: 2
gradient_accumulation_steps: 64
total_train_batch_size: 128
total_eval_batch_size: 2
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: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.6496	0.9926	69	0.7087

Framework versions

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