File size: 1,129 Bytes
8c1a368
 
851e5fd
8c1a368
63d232c
8c1a368
d422439
 
8c1a368
 
 
 
 
d422439
63d232c
d422439
8b4fa69
d422439
8c1a368
d422439
8c1a368
d422439
63d232c
d422439
8c1a368
d422439
8c1a368
d422439
 
 
8c1a368
d422439
 
 
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
---
base_model: eyad-silx/Quasar-3.0-Max
datasets: eyad-silx/Quasar-Max-3.3
library_name: transformers
model_name: Quasar-3.0-Max
tags:
- rl 
- silx
- trl
- sft
licence: license
---

# Quasar Series of Models

<p align="center">
  <img src="https://pbs.twimg.com/media/GlaGzuIWcAAI1JO?format=png&name=small" alt="Quasar Model Image" style="height: 350px;">
</p>

## Introducing Quasar-3.3-Max

This model is provided by **SILX INC**. It has been supervised fine-tuned using the **open-r1** repository. The training data includes sequences of varying lengths (32k, 16k, and 8k) to enhance the model's knowledge and adaptability. 

Quasar-3.3-Max represents the **first step** in the Quasar project before Reinforcement Learning (RL). At this stage, the model's reasoning steps are capped at a **maximum length of 8129 tokens** to optimize processing efficiency and contextual understanding.

Stay tuned for further updates as we advance the Quasar project with RL enhancements!

## Resources
- [Research Paper](https://arxiv.org/abs/2412.06822)
- [Website](https://sicopilot.cloud)

## Founders
- **Eyad Gomaa**
- **Gomaa Salah**