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.

import torch
from transformers import pipeline

model_id = "kyutai/helium-1-2b"

pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

text = 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

Blog post: Helium 1: a modular and multilingual LLM.