README.md

---
base_model: nphearum/psarai-2b
tags:
- transformers
- safetensors
- unsloth
- gemma4
- psarai
- conversational
- multimodal
---

# PsarAI-2B

**PsarAI-2B** is a PsarAI chat model exported in Hugging Face format.

The model uses a Gemma4-style architecture and a PsarAI chat template. The assistant identity in the template is:

> You are PsarAI, created by the PsarAI team under the leadership of an ITC lecturer.

## Files

This repository contains the standard Hugging Face model export:

| File | Purpose |
|---|---|
| `model.safetensors` | model weights |
| `config.json` | model architecture/config |
| `tokenizer.json` | tokenizer |
| `tokenizer_config.json` | tokenizer metadata and special tokens |
| `processor_config.json` | multimodal processor config |
| `chat_template.jinja` | chat formatting template |
| `generation_config.json` | generation defaults |

## Quick Start

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

repo_id = "nphearum/PsarAI-2B"

processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user", "content": "Who created you?"}
]

prompt = processor.tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

inputs = processor.tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
)

print(processor.tokenizer.decode(outputs[0], skip_special_tokens=False))
```

## Chat Template

The template uses Gemma-style tokens:

- `<|turn>system`
- `<|turn>user`
- `<|turn>model`
- `<turn|>`
- `<|channel>thought`
- `<|tool_call>`
- `<|tool_response>`

For normal chatbot use, disable visible thinking when your runtime supports template kwargs:

```python
enable_thinking=False
```

## Suggested Generation Settings

```python
temperature = 0.7
top_p = 0.9
max_new_tokens = 512
```

Use lower temperature, such as `0.2`, for factual or deterministic answers.

## Multimodal Notes

The config includes image, audio, and video processor metadata. Runtime support depends on the installed `transformers` version and model implementation availability.

For GGUF/llama.cpp usage, use the sibling GGUF export repo instead:

```text
nphearum/PsarAI-2B-GGUF
```

## Attribution

Base model metadata in this export is:

```text
nphearum/psarai-2b
```

Keep this metadata for traceability when publishing derived formats.