---
language:
- sv
- "no"
- da
- is
- en
tags:
- text-generation
- swedish
- nordic
- gpt-sw3
- AI-Sweden
- conversational
license: other
library_name: transformers
---

# gpt-sw3-126m-instruct

Smallest GPT-SW3 instruct model (126M parameters). Loads instantly — ideal for testing and prototyping.

**Size:** 126M | **Type:** instruct | **Languages:** Swedish, Norwegian, Danish, Icelandic, English

> Community mirror of [AI-Sweden-Models/gpt-sw3-126m-instruct](https://huggingface.co/AI-Sweden-Models/gpt-sw3-126m-instruct)

---

## Warning and Disclaimer

This model is provided as-is for research and educational purposes.
Community redistribution of AI Sweden's GPT-SW3 under the same modified RAIL license.

**You are responsible for any content you create using this model. Use responsibly.**

The model may reflect biases from training data and may generate inaccurate, offensive,
or inappropriate content. Neither the uploader nor AI Sweden are liable for downstream misuse.
Review the [AI Sweden RAIL license](LICENSE) before any production deployment.

> *"You are responsible for any content you create using this model. Enjoy responsibly."*

---

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "WestCode1357/gpt-sw3-126m-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16)
device = "mps" if torch.backends.mps.is_available() else "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

prompt = "Träd är fina för att"
inputs = tokenizer(prompt, return_tensors="pt").to(device)
out = model.generate(**inputs, max_new_tokens=150, do_sample=True, temperature=0.7)
print(tokenizer.decode(out[0]))
```

### Chat / instruct format

GPT-SW3 instruct uses special tokens. The format is:

```
<|endoftext|><s>User: [your message]<s>Bot: [response]<s>...
```

```python
eos = "<|endoftext|>"
seg = "<s>"
prompt = f"{eos}{seg}User: Vad är huvudstaden i Sverige?{seg}Bot: "
inputs = tokenizer(prompt, return_tensors="pt").to(device)
out = model.generate(
    **inputs, max_new_tokens=200,
    do_sample=True, temperature=0.7, top_p=0.95,
    eos_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=False))
```

## Intended Use

> ⚠️ **These models contain extreme bias and are NOT intended for commercial use.**
> **For scientific and research use only.**

GPT-SW3 was trained on large-scale web data and may reflect harmful societal biases present in that data. It has not been aligned or safety-tuned beyond its original training. Use strictly in controlled research settings. Do not deploy in any consumer-facing or commercial product without thorough evaluation and additional safety measures.

## About GPT-SW3

GPT-SW3 is developed by AI Sweden in collaboration with RISE and WASP WARA for Media and Language.
Trained on 320B tokens: Swedish, Norwegian, Danish, Icelandic, English, and code.

- **Original models:** https://huggingface.co/AI-Sweden-Models
- **Project page:** https://www.ai.se/en/project/gpt-sw3