---
license: gemma
datasets:
- kalomaze/Opus_Instruct_25k
base_model: google/gemma-2-2b-it
---

![image/png](https://cdn-uploads.huggingface.co/production/uploads/660e67afe23148df7ca321a5/AKOTTIQeLVokbsp_Lm7gP.png)

Must put image in repo :3

# Basic info
This is [kalomaze/Opus_Instruct_25k](https://huggingface.co/datasets/kalomaze/Opus_Instruct_25k) over [unsloth/gemma-2-2b-it](https://huggingface.co/unsloth/gemma-2-2b-it)

It saw 39.5M tokens

I have no idea if it's done right but it took 9 hours B)

# Training config:
```
cutoff_len: 1024
dataset: Opus_Instruct_25K
dataset_dir: data
ddp_timeout: 180000000
do_train: true
finetuning_type: lora
flash_attn: auto
fp16: true
gradient_accumulation_steps: 8
include_num_input_tokens_seen: true
learning_rate: 5.0e-05
logging_steps: 5
lora_alpha: 32
lora_dropout: 0
lora_rank: 32
lora_target: all
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 15000
model_name_or_path: unsloth/gemma-2-2b-it
num_train_epochs: 3.0
optim: adamw_8bit
output_dir: saves/Gemma-2-2B-Chat/lora/Final_Opus
packing: false
per_device_train_batch_size: 2
plot_loss: true
preprocessing_num_workers: 16
quantization_bit: 4
quantization_method: bitsandbytes
report_to: none
save_steps: 100
stage: sft
template: gemma
use_unsloth: true
warmup_steps: 0
```