README.md

---
license: apache-2.0
language:
- en
pipeline_tag: text-generation
tags:
- pebble
- language-model
- base-model
- small-language-model
- pytorch
- safetensors
- custom-code
- mamba2
- hybrid
library_name: transformers
---

# Pebble-25M

![Banner](banner.png)

Pebble-25M is a compact, hybrid autoregressive language model. It combines the efficiency of state-space models with the proven performance of attention layers, optimized using a custom Muon + AdamW optimizer split.

## Model Details

- **Architecture:** Hybrid Mamba2 / Transformer
- **Block Pattern:** 3 Mamba2 blocks : 1 Attention block (repeating)
- **Parameters:** ~24,500,000 (25M)
- **Hidden Dimension:** 608
- **Layers:** 8 (6 Mamba2, 2 Attention)
- **Vocab Size:** 2,048 (Custom Byte-Level BPE)
- **Context Length:** 2048
- **Training Tokens:** ~25,000,000,000 (~25 Billion)
- **Optimizer:** Muon (for 2D hidden weights) + AdamW (for embeddings, norms, and scalars)
- **Precision:** fp32 master weights with bf16 autocast

## Dataset Sources

The model was trained on a 25B token subset of the following datasets:

| Dataset | Token Allocation | Share |
|---|---:|---:|
| FineWeb-Edu | 7.50 billion | 30% |
| DCLM | 5.00 billion | 20% |
| Cosmopedia-v2 | 3.75 billion | 15% |
| FineMath-4+ | 3.75 billion | 15% |
| FinePhrase | 3.00 billion | 12% |
| NPset | 2.00 billion | 8% |

## Benchmarks

| Benchmark | **Pebble-25M** | Pebble-25M Chat | Pebble-10M | BananaMind-2-Mini | Random |
|---|---:|---:|---:|---:|---:|
| PIQA | 59.25% | 53.37% | 58.43% | **59.63%** | 50.00% |
| ARC-Easy | 38.17% | 26.68% | 37.29% | **39.86%** | 25.00% |
| ARC-Challenge | 18.60% | 19.62% | 18.60% | **25.68%** | 25.00% |
| HellaSwag | 27.62% | 25.63% | 26.81% | **29.72%** | 25.00% |
| ArithMark-2.0 | 27.60% | 26.20% | **27.64%** | 27.52% | 25.00% |
| ArithMark-3.0 | 33.80% | 28.80% | 32.80% | **34.90%** | 25.00% |

### Evaluation Notes

- PIQA, ARC-Easy, ARC-Challenge, and HellaSwag were evaluated on their respective test splits.
- ArithMark-2.0 was evaluated on its train split due to the lack of a suitable test split.
- ArithMark-3.0 was evaluated on its train split due to the lack of a suitable test split.
- Results were obtained using zero-shot multiple-choice evaluation.
- No task-specific fine-tuning was performed.

## Usage

To run the model for text generation, you will need to install the required dependencies. The included Mamba2 implementation relies on CUDA/Triton kernels and is intended to run on a CUDA-enabled GPU. Ampere-class GPUs or newer are recommended.

> **Note:** The model uses custom architecture code, so you must pass `trust_remote_code=True` when loading both the tokenizer and the model.

### Installation

```bash
pip install transformers huggingface_hub torch
pip install causal-conv1d mamba-ssm  
```

### Generation

Here is a simple Python script to load the model and generate text interactively:

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

MODEL_ID = "basically-ai/Pebble-25M"


def main():
    print("Loading Pebble 25M...")

    tokenizer = AutoTokenizer.from_pretrained(
        MODEL_ID,
        trust_remote_code=True,
    )

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        trust_remote_code=True,
        dtype=torch.float32,
    ).to("cuda")

    model.eval()

    print(
        f"Model loaded successfully! "
        f"VRAM usage: {torch.cuda.memory_allocated() / 1e9:.2f} GB"
    )
    print("Type 'quit' or 'exit' to stop.\n")

    while True:
        prompt = input("You: ")

        if prompt.lower() in ["quit", "exit"]:
            break

        # Tokenize the prompt
        inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

        # Generate text
        print("Pebble: ", end="", flush=True)

        with torch.inference_mode():
            outputs = model.generate(
                **inputs,
                max_new_tokens=100,       # How many tokens to generate
                do_sample=True,           # Use sampling (more creative)
                temperature=0.7,          # Controls randomness
                top_k=50,                 # Consider top 50 tokens
                top_p=0.95,               # Nucleus sampling
                repetition_penalty=1.2,   # Prevent repeating words
            )

        # Decode and print (skip the prompt part)
        generated_text = tokenizer.decode(
            outputs[0][inputs["input_ids"].shape[1]:],
            skip_special_tokens=True,
        )

        print(generated_text)
        print()


if __name__ == "__main__":
    main()
```

## License

Apache 2.0