---
license: gemma
language:
- en
base_model: unsloth/gemma-3-270m-it
datasets:
- Kamisori-daijin/email-datasets-20k
pipeline_tag: text-generation
library_name: transformers
tags:
- sft
- trl
- unsloth
- google
- gemma
- gemma3
- gemma3_text
---
![logo](https://storage.googleapis.com/gweb-developer-goog-blog-assets/images/gemma-3_2.original.png)
A fine-tune of [unsloth/gemma-3-270m-it](https://huggingface.co/unsloth/gemma-3-270m-it) on the [Kamisori-daijin/email-datasets-20k](https://huggingface.co/datasets/Kamisori-daijin/email-datasets-20k) dataset.

## Usage example
**System prompt**
```
You are a helpful assistant.
```
**User prompt**
```
Write a direct and concise business email(under 200 words) from a Founder to a Angry Shareholder regarding notifying about a price change, specifically after winning a major award.
```
**Assistant response**
```
Subject: Regarding Pricing Update - [Company Name]

Dear [Shareholder Name],

I'm writing to address your concerns regarding our recent pricing update. You’re right to raise questions, and I want to be transparent about the reasoning behind it.

We’re incredibly proud to announce that [Company Name] recently won the [Award Name] award – a significant validation of our innovative product and market position. This achievement, coupled with increased customer demand, has necessitated a modest price adjustment to ensure we can continue investing in product development and deliver the exceptional service our customers expect.

We remain confident in the value we provide and believe this change will ultimately benefit all our stakeholders. We welcome the opportunity to discuss this further with you.

Sincerely,
[Name]
Founder, [Company Name]
```
## Model Details
- Base Model: `unsloth/gemma-3-270m-it`
- Parameter Count: 268,098,176
- Precision: torch.bfloat16

## Hardware
- GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
- Announced: Mar 17th, 2025
- Release Date: Mar 18th, 2025
- Memory Type: GDDR7
- Bandwidth: 1.79 TB/s
- Memory Size: 96 GB
- Memory Bus: 512 bit
- Shading Units: 24064
- TDP: 600W

## Training Settings
### PEFT
- Rank: 32
- LoRA alpha: 64
- Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- Gradient checkpointing: unsloth

### SFT
- Epoch: 4
- Batch size: 32
- Gradient Accumulation steps: 1
- Warmup ratio: 0.05
- Learning rate: 0.0002
- Optimizer: adamw_torch_fused
- Learning rate scheduler: cosine

## Training stats
- Date: 2026-03-25T12:51:43.831886
- Peak VRAM usage: 16.834 GB
- Global step: 2360
- Training runtime (seconds): 470.094
- Average training loss: 1.2040837437419567
- Final validation loss: 1.2054944038391113

## Framework versions
- Unsloth: 2026.3.11
- TRL: 0.22.2
- Transformers: 4.56.2
- Pytorch: 2.10.0+cu128
- Datasets: 4.8.4
- Tokenizers: 0.22.2

## License
This model is released under the Gemma license. See the [Gemma Terms of Use](https://ai.google.dev/gemma/terms) and [Prohibited Use Policy](https://policies.google.com/terms/generative-ai/use-policy) regarding the use of Gemma-generated content.