---
base_model: google/gemma-4-e2b-it
library_name: transformers
pipeline_tag: image-text-to-text
---
# gemma4-e2b-webvid4K_FT
Full fine-tune of `google/gemma-4-e2b-it` on AI-generated video data derived from WebVid.
## Training
- Dataset: `bear7011/gemma-4-e4b-webvid-4K`
- Samples: 3,941 video instruction examples
- Method: full fine-tuning, no LoRA
- Precision: bfloat16
- GPUs: 4
- DeepSpeed: ZeRO-3 with CPU optimizer and parameter offload
- Epochs: 1
- Global steps: 124
- Per-device batch size: 1
- Gradient accumulation steps: 8
- Optimizer: AdamW
- Learning rate: 5e-6
- Projector learning rate: 5e-6
- Image encoder learning rate: 0.0
- Weight decay: 0.01
- Warmup ratio: 0.03
- LR scheduler: cosine
- Gradient checkpointing: enabled
- Max sequence length: 2304
- Final training loss: 1.9510
Checkpoints and training logs are not included in this repository.