100 lines
4.8 KiB
Markdown
100 lines
4.8 KiB
Markdown
---
|
||
license: apache-2.0
|
||
datasets:
|
||
- prithivMLmods/Deepthink-Reasoning-Tamil
|
||
language:
|
||
- ta
|
||
- en
|
||
base_model:
|
||
- Qwen/Qwen2.5-3B-Instruct
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
tags:
|
||
- text-generation-inference
|
||
- Qwen
|
||
---
|
||
# **Qwen2.5-3B-Tamil-Exp**
|
||
|
||
**Qwen2.5-3B-Tamil-Exp** is built on the robust Qwen2.5 architecture and has been specifically adapted to excel at Tamil language tasks. By incorporating training log entries from the prithivMLmods/Deepthink-Reasoning-Tamil dataset along with the proven reasoning framework of Qwen models, this 3B-parameter variant achieves enhanced chain-of-thought reasoning and logical problem solving—especially tailored for Tamil. Its improvements extend to context understanding, structured data processing, and long-context comprehension, making it ideal for complex reasoning tasks, instruction-following, and text generation in Tamil and other languages.
|
||
|
||
### **Key Improvements**
|
||
1. **Advanced Reasoning & Logic:**
|
||
Optimized for multi-step problem solving and logical deduction. Fine-tuning on the Deepthink-Reasoning-Tamil entries further refines its reasoning capabilities in Tamil contexts.
|
||
|
||
2. **Fine-Tuned Instruction Following:**
|
||
Generates precise responses and structured outputs (such as JSON), making it well-suited for dialog-based applications and code generation tasks that require strict adherence to Tamil language instructions.
|
||
|
||
3. **Greater Adaptability:**
|
||
Excels in role-playing scenarios, multi-turn dialogues, and diverse system prompts with a focus on culturally nuanced Tamil content while maintaining support for multiple languages.
|
||
|
||
4. **Long-Context Support:**
|
||
Capable of handling extended inputs (up to 64K tokens) and generating outputs of up to 4K tokens, enabling the processing of detailed and lengthy Tamil texts.
|
||
|
||
5. **Multilingual Proficiency with Tamil Focus:**
|
||
While supporting over 20 languages, the model’s training emphasis on Tamil ensures superior performance on tasks involving Tamil language understanding and generation.
|
||
|
||
### **Intended Use**
|
||
- **Advanced Logical & Analytical Reasoning:**
|
||
Ideal for solving multi-step problems and deductive reasoning tasks, especially those presented in Tamil.
|
||
- **Mathematical & Scientific Computation:**
|
||
Supports theorem proving, complex calculations, and retrieval of scientific knowledge with an emphasis on Tamil terminology.
|
||
- **Code Generation & Debugging:**
|
||
Generates optimized code, detects errors, and enhances programming workflows with support for Tamil documentation or comments.
|
||
- **Structured Data Analysis:**
|
||
Processes tables, JSON, and other structured formats, which is particularly useful for localized applications requiring Tamil language outputs.
|
||
- **Multilingual Reasoning & Translation:**
|
||
While excelling in Tamil, it is also proficient in other languages for international applications.
|
||
- **Extended Text Generation:**
|
||
Capable of producing research papers, instructional guides, and in-depth reports in Tamil.
|
||
|
||
### **Quickstart with Transformers**
|
||
|
||
Below is an example of how to load and use the model with the Hugging Face Transformers library:
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_name = "your_org/Qwen2.5-3B-Tamil-Exp"
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
||
prompt = "தமிழில் தர்க்கரீதியான எண்ணத்தை விளக்குங்கள்." # "Explain the concept of logical reasoning in Tamil."
|
||
messages = [
|
||
{"role": "system", "content": "நீங்கள் ஒரு தமிழில் சிறந்த தர்க்கரீதியான உதவியாளர்."},
|
||
{"role": "user", "content": prompt}
|
||
]
|
||
text = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=False,
|
||
add_generation_prompt=True
|
||
)
|
||
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
|
||
generated_ids = model.generate(
|
||
**model_inputs,
|
||
max_new_tokens=256
|
||
)
|
||
generated_ids = [
|
||
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
|
||
]
|
||
|
||
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
||
print(response)
|
||
```
|
||
|
||
### **Limitations**
|
||
1. **Moderate Computational Requirements:**
|
||
Requires mid-end consumer GPUs for optimal inference.
|
||
2. **Language-Specific Variability:**
|
||
While performance is strong for Tamil, results may vary for other supported languages.
|
||
3. **Potential Error Accumulation:**
|
||
Extended outputs may sometimes introduce inconsistencies.
|
||
4. **Limited Real-World Awareness:**
|
||
The model’s knowledge is based on its training data and may not include recent events.
|
||
5. **Prompt Sensitivity:**
|
||
High-quality responses depend on the clarity and specificity of the input prompt. |