See axolotl config

axolotl version: 0.10.0.dev0

adapter: lora
base_model: Qwen/Qwen2.5-7B
bf16: auto
dataset_prepared_path: text2svg-prepared
datasets:
- chat_template: tokenizer_default
  field_messages: messages
  path: thesantatitan/text2svg-stack-follow-constraints-10k
  roles_to_train:
  - assistant
  split: train
  type: chat_template
debug: null
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
early_stopping_patience: null
eval_sample_packing: false
flash_attention: true
fp16: false
fsdp: null
fsdp_config: null
gradient_accumulation_steps: 16
gradient_checkpointing: true
group_by_length: false
hub_model_id: thesantatitan/bagel-svg-sft-new-rank128
hub_strategy: every_save
learning_rate: 0.0001
load_in_4bit: false
load_in_8bit: false
local_rank: null
logging_steps: 1
lora_alpha: 32
lora_dropout: 0.05
lora_fan_in_fan_out: null
lora_model_dir: null
lora_modules_to_save:
- embed_tokens
- lm_head
lora_r: 128
lora_target_linear: true
lr_scheduler: cosine
micro_batch_size: 1
model_type: AutoModelForCausalLM
num_epochs: 1
optimizer: adamw_torch
output_dir: ./lora-out
pad_to_sequence_len: false
resume_from_checkpoint: null
sample_packing: false
save_steps: 20
sequence_len: 4096
strict: false
tf32: false
tokenizer_type: AutoTokenizer
train_on_inputs: false
trust_remote_code: true
val_set_size: 0.05
wandb_entity: null
wandb_project: svg-sft-bagel-saved
wandb_run_id: sexyrunbagel1
wandb_watch: null
warmup_steps: 10
weight_decay: 0.0
xformers_attention: null

bagel-svg-sft-new-rank128

This model is a fine-tuned version of Qwen/Qwen2.5-7B on the thesantatitan/text2svg-stack-follow-constraints-10k dataset. It achieves the following results on the evaluation set:

• Loss: 0.7283

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: 16
• total_train_batch_size: 64
• 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: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.6641	1.0	139	0.7283

Framework versions

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