Commit ·
bf3c992
1
Parent(s): e1e2cee
Update README.md
Browse files
README.md
CHANGED
|
@@ -11,28 +11,99 @@ license: cc-by-sa-4.0
|
|
| 11 |
|
| 12 |
This page houses `Helium 6B` models trained using either a sequential pretraining on temporally ordered data or using a standard pretraining on shuffled data. The architecture is derived from [Helium 2B](https://huggingface.co/kyutai/helium-1-2b).
|
| 13 |
|
| 14 |
-
## Models Details
|
| 15 |
|
|
|
|
|
|
|
| 16 |
|
|
|
|
| 17 |
|
| 18 |
-
|
| 19 |
|
| 20 |
-
|
| 21 |
-
### Licensing
|
| 22 |
|
| 23 |
-
|
| 24 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
-
|
| 27 |
|
|
|
|
|
|
|
|
|
|
| 28 |
|
|
|
|
| 29 |
|
| 30 |
-
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
```
|
| 33 |
|
| 34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
## Citations
|
| 37 |
|
| 38 |
If you use one of these models, please cite:
|
|
|
|
| 11 |
|
| 12 |
This page houses `Helium 6B` models trained using either a sequential pretraining on temporally ordered data or using a standard pretraining on shuffled data. The architecture is derived from [Helium 2B](https://huggingface.co/kyutai/helium-1-2b).
|
| 13 |
|
|
|
|
| 14 |
|
| 15 |
+
- **Developed by:** Kyutai
|
| 16 |
+
- **Model type:** Large Language Model
|
| 17 |
|
| 18 |
+
## Uses
|
| 19 |
|
| 20 |
+
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
| 21 |
|
| 22 |
+
### Direct Use
|
|
|
|
| 23 |
|
| 24 |
+
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
|
| 25 |
+
|
| 26 |
+
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.
|
| 27 |
+
For most downstream use cases, the model should be aligned with supervised fine-tuning, RLHF or related methods.
|
| 28 |
+
|
| 29 |
+
### Out-of-Scope Use
|
| 30 |
+
|
| 31 |
+
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
|
| 32 |
+
|
| 33 |
+
The model should not be used in other languages than the ones on which it was trained.
|
| 34 |
+
The model is not intended to be used for any malicious or illegal activities of any kind.
|
| 35 |
+
The model was not fine-tuned to follow instructions, and thus should not be used as such.
|
| 36 |
|
| 37 |
+
## Bias, Risks, and Limitations
|
| 38 |
|
| 39 |
+
Helium-1 is a base language model, which was not aligned to human preferences.
|
| 40 |
+
As such, the model can generate incorrect, biased, harmful or generally unhelpful content.
|
| 41 |
+
Thus, the model should not be used for downstream applications without further alignment, evaluations and mitigations of risks.
|
| 42 |
|
| 43 |
+
## How to Get Started with the Model
|
| 44 |
|
| 45 |
+
Use the code below to get started with the model.
|
| 46 |
|
| 47 |
+
```python
|
| 48 |
+
import torch
|
| 49 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 50 |
+
|
| 51 |
+
model_id = "kyutai/Sequential_Helium_6B"
|
| 52 |
+
|
| 53 |
+
model = AutoModelForCausalLM.from_pretrained(model_id).cuda()
|
| 54 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 55 |
```
|
| 56 |
|
| 57 |
+
To load a specific checkpoint, e.g. the last checkpoint from the sequential pretraining (cool-downed) before any 2025 data:
|
| 58 |
+
|
| 59 |
+
```python
|
| 60 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, subfolder='sequential_2024').cuda()
|
| 61 |
+
```
|
| 62 |
+
## Training Details
|
| 63 |
+
|
| 64 |
+
### Training Data
|
| 65 |
+
|
| 66 |
+
Helium-6B checkpoints were trained on data from Common Crawl, which was preprocessed with the [dactory](https://github.com/kyutai-labs/dactory) library.
|
| 67 |
+
|
| 68 |
+
|
| 69 |
|
| 70 |
+
## Evaluation
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
#### Testing Data
|
| 76 |
+
|
| 77 |
+
The model was evaluated using [OLMES](https://arxiv.org/abs/2406.08446) a LLM evaluation benchmark based on, MMLU, ARC Easy & Challenge, Open Book QA, Common Sense QA, Physical Interaction QA, Social Interaction QA, HellaSwag, WinoGrande and BoolQA.
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
#### English Results
|
| 82 |
+
|
| 83 |
+
| Benchmark | Sequential-Helium-6B | Shuffled-Helium-6B (2.5T tokens) |
|
| 84 |
+
|--------------|:------:|:------:|
|
| 85 |
+
| | | | | | |
|
| 86 |
+
| MMLU | 58.8 | 56.4 |
|
| 87 |
+
| ARC E | 87.6 | 86.7 |
|
| 88 |
+
| ARC C | 74.5 | 72.1 |
|
| 89 |
+
| OBQA | 72.8 | 73.2 |
|
| 90 |
+
| CSQA | 73.1 | 74.3 |
|
| 91 |
+
| PIQA | 80.3 | 80.2 |
|
| 92 |
+
| SIQA | 67.0 | 66.2 |
|
| 93 |
+
| HS | 79.1 | 81.3 |
|
| 94 |
+
| WG | 73.0 | 73.1 |
|
| 95 |
+
| BoolQA | 83.9 | 83.9 |
|
| 96 |
+
| | | |
|
| 97 |
+
| OLMES | 75.0 | 74.7 |
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
### Uses
|
| 101 |
+
|
| 102 |
+
As described in the [paper](),
|
| 103 |
+
### Licensing
|
| 104 |
+
|
| 105 |
+
Helium 6B models are licensed under the CC-BY-SA 4.0 license.
|
| 106 |
+
|
| 107 |
## Citations
|
| 108 |
|
| 109 |
If you use one of these models, please cite:
|