81 lines
4.2 KiB
Markdown
81 lines
4.2 KiB
Markdown
---
|
|
library_name: transformers
|
|
tags:
|
|
- math
|
|
- text-generation-inference
|
|
- code
|
|
- 3B
|
|
license: llama3.2
|
|
language:
|
|
- en
|
|
base_model:
|
|
- meta-llama/Llama-3.2-3B-Instruct
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
|
|
# **Ganymede-Llama-3.3-3B-Preview**
|
|
|
|
> Ganymede-Llama-3.3-3B-Preview is based on the **Llama-3.2-3B-Instruct** architecture, featuring **unlocked abliterated** capabilities and **improved mathematical analysis**. Fine-tuned on a high-quality synthetic dataset derived from Llama's **Instruct** series, it excels in **chain-of-thought (CoT) reasoning, logical problem-solving, and structured data comprehension**. The model is ideal for complex reasoning tasks, instruction-following, and text generation, with superior adaptability across **multi-turn conversations and long-context tasks**.
|
|
|
|
### **Key Improvements**
|
|
1. **Unlocked Abliterated Reasoning**: Enhanced **multi-step problem-solving, logical deduction, and contextual analysis**.
|
|
2. **Mathematical & Analytical Excellence**: Stronger capabilities in **math problem-solving, theorem proving, and complex numerical analysis**.
|
|
3. **Fine-Tuned Instruction Following**: Generates structured responses (e.g., **JSON, XML, Markdown**) and **long-form text (4K+ tokens)**.
|
|
4. **Extended Long-Context Support**: Handles up to **128K tokens** with improved memory retention and coherence over long passages.
|
|
5. **Advanced Adaptability**: Excels in **role-playing, multi-turn dialogues, and diverse system prompts**.
|
|
6. **Multilingual Proficiency**: Supports over **20 languages**, including **English, Chinese, French, Spanish, Portuguese, German, and more**.
|
|
|
|
### **Quickstart with Transformers**
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_name = "prithivMLmods/Ganymede-Llama-3.3-3B-Preview"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
torch_dtype="auto",
|
|
device_map="auto",
|
|
trust_remote_code=True
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
prompt = "Explain the concept of logical reasoning in AI."
|
|
messages = [
|
|
{"role": "system", "content": "You are an expert AI assistant specialized in reasoning, logic, and mathematical analysis."},
|
|
{"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)
|
|
```
|
|
|
|
### **Intended Use**
|
|
- **Advanced Logical & Analytical Reasoning**: Designed for **multi-step problem-solving, deductive reasoning, and cognitive tasks**.
|
|
- **Enhanced Mathematical Computation**: Excels in **numerical analysis, theorem proving, symbolic reasoning, and complex calculations**.
|
|
- **Code Generation & Debugging**: Generates **optimized code**, detects **bugs**, and enhances **programming workflows**.
|
|
- **Structured Data Processing**: Handles **tables, JSON, and structured formats** for data-centric applications.
|
|
- **Multilingual Reasoning & Translation**: High proficiency across **20+ languages** for **global AI applications**.
|
|
- **Extended Text Generation**: Ideal for generating **technical documentation, research papers, instructional guides, and in-depth reports**.
|
|
|
|
### **Limitations**
|
|
1. **Moderate Computational Requirements**: Requires **mid-to-high-end consumer GPUs** for optimal inference.
|
|
2. **Language-Specific Variability**: Performance may differ across languages, particularly for **low-resource languages**.
|
|
3. **Potential Error Accumulation**: Long-form text generation may introduce **inconsistencies** over extended outputs.
|
|
4. **Limited Real-World Awareness**: Knowledge is restricted to **training data** and may not reflect **recent world events**.
|
|
5. **Prompt Sensitivity**: The quality of responses depends on **prompt clarity and specificity**. |