• Model creator: https://huggingface.co/flydust/
• Original model: https://huggingface.co/flydust/gemma-2-2b-magpie-gemma2-9b-flash/

Original model description:

• axolotl
• generated_from_trainer model-index:
• name: gemma-2-2b-magpie-gemma2-9b-flash results: []

---

See axolotl config

axolotl version: 0.4.1

yaml

1base_model: google/gemma-2-2b 2model_type: Gemma2ForCausalLM 3tokenizer_type: AutoTokenizer 4chat_template: gemma 5 6load_in_8bit: false 7load_in_4bit: false 8strict: false 9 10datasets: 11 - path: flydust/Magpie-100k-Gemma2-9B 12 type: sharegpt 13 chat_template: gemma 14dataset_prepared_path: last_run_prepared 15val_set_size: 0.001 16output_dir: axolotl_out/gemma-2-2b-magpie-gemma2-9b-flash 17 18sequence_len: 4096 19sample_packing: true 20eval_sample_packing: false 21pad_to_sequence_len: true 22 23wandb_project: SynDa 24wandb_entity: 25wandb_watch: 26wandb_name: gemma-2-2b-magpie-gemma2-9b-flash 27wandb_log_model: 28hub_model_id: flydust/gemma-2-2b-magpie-gemma2-9b-flash 29 30gradient_accumulation_steps: 8 31micro_batch_size: 1 32num_epochs: 2 33optimizer: paged_adamw_8bit 34lr_scheduler: cosine 35learning_rate: 2e-5 36 37train_on_inputs: false 38group_by_length: false 39bf16: true 40fp16: 41tf32: false 42 43gradient_checkpointing: true 44gradient_checkpointing_kwargs: 45 use_reentrant: false 46early_stopping_patience: 47resume_from_checkpoint: 48logging_steps: 1 49xformers_attention: 50# Disable flash attention 51flash_attention: true 52# sdp_attention: falses 53# eager_attention: true 54 55warmup_ratio: 0.1 56evals_per_epoch: 5 57eval_table_size: 58saves_per_epoch: 1 59debug: 60deepspeed: 61weight_decay: 0.0 62fsdp: 63fsdp_config: 64special_tokens: 65

gemma-2-2b-magpie-gemma2-9b-flash

• Loss: 0.6632

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 86
• num_epochs: 2

Training results

Framework versions

• Transformers 4.45.0.dev0
• Pytorch 2.4.1+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1