---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- causal-lm
- base-model
- small-language-model
- custom_code
- muon
- hummingbird
- hummingbird-v2
datasets:
- HuggingFaceFW/fineweb-edu
- HuggingFaceTB/smollm-corpus
- HuggingFaceTB/finemath
- mlfoundations/dclm-baseline-1.0
- allenai/dolma3_dolmino_mix-100B-1025
- allenai/dolma3_dolmino_mix-10B-1025
- HuggingFaceFW/finepdfs-edu
---

![Hummingbird-V2 banner with daisies](hummingbird-v2-banner.jpg)

# Hummingbird-V2

Hummingbird-V2 is a 9.6M-parameter English base language model trained from scratch on
10 billion tokens. It is designed for text completion and continuation scoring.

## Model series

Hummingbird-V2 follows [Hummingbird-V1](https://huggingface.co/juinron/Hummingbird-V1)
in the same small language model series. V2 was trained from a fresh initialization.

## Architecture

| Property | Value |
|---|---:|
| Parameters | 9,592,720 |
| Layers / hidden size | 14 / 240 |
| Vocabulary | 4,096-token digit-aware byte-level BPE |
| Maximum context | 2,048 tokens |
| Architecture | Decoder-only Transformer with grouped-query attention and SwiGLU |

## Training and data

The model was trained on this balanced mix:

| Source | Share |
|---|---:|
| FineWeb-Edu sources | 55% |
| Cosmopedia v2 | 15% |
| FineMath 4+ | 10% |
| DCLM baseline | 10% |
| Dolma 3 science, question answering, and code | 7% |
| FinePDFs-Edu | 3% |

Training used Muon and AdamW with a 512-token context. See
[training data](TRAINING_DATA.md) for source details and
[training provenance](training/provenance.json) for the full recipe.

## Zero-shot evaluation

Author-run zero-shot results for the 10B-token checkpoint. Task scores are normalized
continuation accuracy in percent; the Intelligence Index is a chance-normalized composite.

| Benchmark | Score |
|---|---:|
| HellaSwag | 27.63 |
| ARC-Easy | 39.39 |
| ARC-Challenge | 21.16 |
| PIQA | 57.40 |
| ArithMark-3 | 36.00 |
| Chance-normalized Intelligence Index | **9.556** |

These evaluations helped select the released checkpoint; results have not been
independently verified.

## Use

Load `juinron/Hummingbird-V2` with Transformers and the packaged custom code:

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

model_id = "juinron/Hummingbird-V2"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True).eval()

inputs = tokenizer("The color of the sky is", return_tensors="pt")
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=32, do_sample=False, use_cache=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

`trust_remote_code=True` loads the packaged model code. Review it before use.

## Limitations

Hummingbird-V2 is small and English-focused. It can make factual or reasoning errors and
is not instruction-tuned or safety-aligned. Do not rely on it for consequential decisions.

## License

The model package is released under [Apache-2.0](LICENSE). Third-party dataset notices
are included in [NOTICE](NOTICE) and [TRAINING_DATA.md](TRAINING_DATA.md).