File size: 1,393 Bytes
6024e15
 
 
 
206899b
 
3909b06
706a33e
206899b
5b05bb0
 
d7c6bda
5b05bb0
206899b
 
3909b06
206899b
e917fa2
 
 
 
 
 
 
 
206899b
 
3909b06
206899b
60561a7
 
 
afc1c63
60561a7
206899b
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
---
license: apache-2.0
language:
- en
---

# Model Card for TurboSparse-Mixtral
The [TurboSparse-Mixtral](https://arxiv.org/abs/2406.05955) Large Language Model (LLM) is an sparsified version of the Mixtral.

<img src="takeaway.png" alt="avatar" width="300" height="200"/>

The average performance is evaluated using benchmarks from the OpenLLM Leaderboard.

## Inference

Our code for accelerating TurboSparse-Mixtral is currently being refined. Stay tuned! Now you can run this model like dense model.

## Chat-Template

During sparsification, we also utilize some SFT datasets.
We take ChatML as our chat template:
```
<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n
```

## Allow Finetuning

As we merged the predictors for FFN neurons in models, you can finetune TurboSparse-Mixtral with any framework and algorithm.

## Limitations
* TurboSparse, having just undergone training with 150B tokens, may still exhibit performance gaps in certain tasks.
* The TurboSparse model has only been trained on English-language datasets, hence its capabilities in other languages are still lacking.
* The model may produce unexpected outputs due to its small size, limited training tokens and probabilistic generation paradigm.
  
## License

The model is licensed under Apache-2.0, while model weights are fully open for academic research and also allow **free** commercial usage.