README.md

---
language: [pl]
license: llama3.1
pipeline_tag: text-generation
library_name: transformers
tags:
  - llama
  - llama-3.1
  - polish
  - grpo
  - reasoning
  - safetensors
datasets:
  - openai/gsm8k
base_model: CYFRAGOVPL/Llama-PLLuM-8B-instruct
base_model_relation: finetune
---

# Llama-PLLuM-8B-instruct-ArtexIT-reasoning

**Built with Llama**

This repository contains a GRPO fine‑tune of [`CYFRAGOVPL/Llama-PLLuM-8B-instruct`] trained on **GSM8K** (MIT).
We publish both **Hugging Face (safetensors)** and **GGUF** artifacts (Q8_0, Q5_K_M) for use with `llama.cpp`.


## What is this?
- **Base**: Meta Llama 3.1 → PLLuM 8B Instruct (Polish) → GRPO fine‑tune (math / word problems).
- **Context**: ~131k (based on GGUF header).
- **Message format**: Llama `[INST] ... [/INST]` + explicit reasoning / answer tags (see below).
- **Default chat template**: The tokenizer includes a default system instruction enforcing the two‑block format.


## Prompt format

The model expects Llama chat formatting and supports explicit tags:

- **Reasoning**: `<think> ... </think>`  
- **Final answer**: `<answer> ... </answer>`

**Example**
```text
[INST] Rozwiąż: 12 * 13 = ? [/INST]
<think>12*13 = 156.</think>
<answer>156</answer>
```

## Quickstart

### Transformers (PyTorch)

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

repo = "ARTEXIT/Llama-PLLuM-8B-instruct-ArtexIT-reasoning"
tok = AutoTokenizer.from_pretrained(repo, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")

prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Podaj 3 miasta w Polsce."}],
    add_generation_prompt=True,
    tokenize=False,
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=64)
print(tok.decode(out[0], skip_special_tokens=False))
```


## Training (brief)

- **Method**: GRPO (policy‑gradient reinforcement learning with multiple reward functions).
- **Data**: `openai/gsm8k` — License: **MIT**.
- **Goal**: consistent two‑block outputs (reasoning + final answer) using the training tags.


## License & Attribution

This repository contains derivatives of **Llama 3.1** and **PLLuM**:

- **Llama 3.1 Community License** applies. When redistributing, you must:
  - include a copy of the license and **prominently display “Built with Llama”**,
  - include **“Llama” at the beginning of any distributed model’s name** if it was created, trained or fine‑tuned using Llama materials,
  - keep a **NOTICE** file with the following line:  
    `Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.`
  - comply with the **Acceptable Use Policy (AUP)**.
- **PLLuM**: please cite the PLLuM work (see **Citation** below).
- **Data**: GSM8K is MIT‑licensed; include dataset attribution.

This repo includes:
- `LICENSE` — full text of the **Llama 3.1 Community License**
- `USE_POLICY.md` — pointer to the official **Acceptable Use Policy**
- `NOTICE` — required Llama attribution line

> If your (or your affiliates’) products exceeded **700M monthly active users** on the Llama 3.1 release date, you must obtain a separate license from Meta before exercising the rights in the Llama 3.1 license.


## Citation

If you use PLLuM in research or deployments, please cite:

```bibtex
@unpublished{pllum2025,
    title={PLLuM: A Family of Polish Large Language Models},
    author={PLLuM Consortium},
    year={2025}
}
```