Helium-1-2b
Model Description
Helium-1 is a lightweight language model with 2B parameters, targeting edge and mobile devices.
It supports the 24 official languages of the European Union.
⚠️ Helium-1 is a base model, which was not fine-tuned to follow instructions or human preferences.
For most downstream use cases, the model should be aligned with supervised fine-tuning, RLHF or related methods.
- Developed by: Kyutai
- Model type: Large Language Model
- Language(s) (NLP): Bulgarian, Czech, Danish, German, Greek, English, Spanish, Estonian, Finnish, French, Irish, Croatian, Hungarian, Italian, Lithuanian, Latvian, Maltese, Dutch, Polish, Portuguese, Romanian, Slovak, Slovenian, Swedish.
- License: CC-BY-SA 4.0
- Terms of use: As a model distilled from Gemma 2, Helium 1 is subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms
Uses
Direct Use
The intended use of the Helium model is research and development of natural language processing systems, including but not limited to language generation and understanding.
The model can be used in Bulgarian, Czech, Danish, German, Greek, English, Spanish, Estonian, Finnish, French, Irish, Croatian, Hungarian, Italian, Lithuanian, Latvian, Maltese, Dutch, Polish, Portuguese, Romanian, Slovak, Slovenian, Swedish.
For most downstream use cases, the model should be aligned with supervised fine-tuning, RLHF or related methods.
Out-of-Scope Use
The model should not be used in other languages than the ones on which it was trained.
The model is not intended to be used for any malicious or illegal activities of any kind.
The model was not fine-tuned to follow instructions, and thus should not be used as such.
Bias, Risks, and Limitations
Helium-1 is a base language model, which was not aligned to human preferences.
As such, the model can generate incorrect, biased, harmful or generally unhelpful content.
Thus, the model should not be used for downstream applications without further alignment, evaluations and mitigations of risks.
How to Get Started with the Model
Use the code below to get started with the model.
1import torch
2from transformers import pipeline
3
4model_id = "kyutai/helium-1-2b"
5
6pipe = pipeline(
7 "text-generation",
8 model=model_id,
9 torch_dtype=torch.bfloat16,
10 device_map="auto"
11)
12
13text = pipe("Hello, today is a great day to")
Training Details
Training Data
Helium-1 was trained on data from Common Crawl, which was preprocessed with the dactory library.
Evaluation
Testing Data
The model was evaluated on MMLU, TriviaQA, NaturalQuestions, ARC Easy & Challenge, Open Book QA, Common Sense QA,
Physical Interaction QA, Social Interaction QA, HellaSwag, WinoGrande, Multilingual Knowledge QA, FLORES 200.
Metrics
We report accuracy on MMLU, ARC, OBQA, CSQA, PIQA, SIQA, HellaSwag, WinoGrande.
We report exact match on TriviaQA, NQ and MKQA.
We report BLEU on FLORES.
English Results
| Benchmark | Helium-1 | HF SmolLM2 (1.7B) | Gemma-2 (2.6B) | Llama-3.2 (3B) | Qwen2.5 (1.5B) |
|---|
| | | | | |
| MMLU | 52.0 | 50.4 | 53.1 | 56.6 | 61.0 |
| NQ | 16.5 | 15.1 | 17.7 | 22.0 | 13.1 |
| TQA | 46.5 | 45.4 | 49.9 | 53.6 | 35.9 |
| ARC E | 82.2 | 81.8 | 81.1 | 84.6 | 89.7 |
| ARC C | 64.6 | 64.7 | 66.0 | 69.0 | 77.2 |
| OBQA | 65.4 | 61.4 | 64.6 | 68.4 | 73.8 |
| CSQA | 63.6 | 59.0 | 64.4 | 65.4 | 72.4 |
| PIQA | 78.5 | 77.7 | 79.8 | 78.9 | 76.0 |
| SIQA | 62.3 | 57.5 | 61.9 | 63.8 | 68.7 |
| HS | 73.6 | 73.2 | 74.7 | 76.9 | 67.5 |
| WG | 66.9 | 65.6 | 71.2 | 72.0 | 64.8 |
| | | | | |
| Average | 61.1 | 59.3 | 62.2 | 64.7 | 63.6 |
Multilingual Results
| Benchmark | Helium-1 | Gemma-2 (2.6B) | Llama-3.2 (3B) |
|---|
| | | |
| ARC E | 71.1 | 65.8 | 68.2 |
| ARC C | 54.8 | 51.1 | 52.6 |
| MMLU | 44.8 | 43.1 | 45.3 |
| HS | 51.9 | 49.9 | 48.4 |
| FLORES | 20.6 | 21.9 | 19.8 |
| MKQA | 16.5 | 17.2 | 19.7 |
| | | |
| Average | 43.3 | 41.5 | 42.3 |
Technical Specifications
Model Architecture and Objective
| Hyperparameter | Value |
|---|
| Model dimension | 2048 |
| MLP dimension | 8192 |
| Layers | 28 |
| Heads | 16 |
| RoPE theta | 20,000 |
| Context size | 4096 |
| Max learning rate | 2.4e-04 |
| Total steps | 500,000 |
| Weight decay | 0.1 |
| Gradient clip | 1.0 |
Hardware
The model was trained on 64 NVIDIA H100 Tensor Core GPUs.
Software
The model was trained using Jax.
Citation
