Imran1 commited on
Commit
25621ed
·
verified ·
1 Parent(s): f027917

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +26 -1
README.md CHANGED
@@ -1,3 +1,28 @@
1
  ---
2
  license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ ---
4
+
5
+ # QWEN2.5-32B-2600s-FP8: Advanced Multilingual Translation Model
6
+
7
+ ## Overview
8
+ **Imran1/QWEN2.5-32B-Translation** is a fine-tuned version of Qwen 2.5 32B, specifically optimized for multilingual translation across **16 different languages**. This model has been extensively fine-tuned to enhance its translation capabilities, making it competitive with high-tier models like 72B in terms of translation accuracy and fluency.
9
+
10
+ ## Fine-Tuning Process
11
+ ### Data Collection
12
+ To improve the model's understanding and translation capabilities, we curated and synthesized a large dataset consisting of:
13
+ - High-quality multilingual conversational datasets.
14
+ - Real-world dialogues spanning general, business, and technical domains.
15
+ - Translated datasets covering diverse linguistic structures and idiomatic expressions.
16
+
17
+ ### Multilingual Enhancement
18
+ To advance its translation capabilities, we leveraged:
19
+ - **Translation Expansion**: The collected dataset was translated into **16 different languages** to ensure robust multilingual performance.
20
+ - **Benchmarking Against High-Tier Models**: We utilized state-of-the-art translation models, including **Gemini** and other top-ranking translation models with high BLEU and COMET scores, to refine our translation quality.
21
+ - **Reinforcement Learning with Human Feedback (RLHF)**: Translation outputs were evaluated and iteratively improved based on feedback from native speakers and linguistic experts.
22
+
23
+ ### Training and Optimization
24
+ - **Base Model**: Qwen 2.5 32B FP8
25
+ - **Fine-Tuning Framework**: LoRA + QLoRA for efficient training
26
+ - **Batch Size**: Optimized for multi-GPU environments
27
+ - **Precision**: FP8 for efficient computation without sacrificing performance
28
+ - **Training Iterations**: Over 2600 steps on **multi-H100 GPUs**