120 lines
4.4 KiB
Markdown
120 lines
4.4 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
datasets:
|
|||
|
|
- open-r1/Mixture-of-Thoughts
|
|||
|
|
- nvidia/OpenCodeReasoning
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
base_model:
|
|||
|
|
- prithivMLmods/Qwen3-1.7B-ft-bf16
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library_name: transformers
|
|||
|
|
tags:
|
|||
|
|
- text-generation-inference
|
|||
|
|
- math
|
|||
|
|
- science
|
|||
|
|
- moe
|
|||
|
|
- code
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
# **Capricornus-MoT-1.7B-Supreme1**
|
|||
|
|
|
|||
|
|
> **Capricornus-MoT-1.7B-Supreme1** is a **high-precision, multi-domain expert model** fine-tuned from **Qwen3-1.7B**, built for **code generation**, **mathematical reasoning**, **scientific analysis**, and **open technical inference**. Trained on the **Mixture of Thoughts (MoT)** dataset with combined expert clusters in **code, math, and science**, and enhanced with an **Open Code Reasoning** dataset, it delivers powerful symbolic and structured outputs in a wide range of STEM and reasoning domains.
|
|||
|
|
|
|||
|
|
> \[!note]
|
|||
|
|
> GGUF: [https://huggingface.co/prithivMLmods/Capricornus-MoT-1.7B-Supreme1-GGUF](https://huggingface.co/prithivMLmods/Capricornus-MoT-1.7B-Supreme1-GGUF)
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## **Key Features**
|
|||
|
|
|
|||
|
|
1. **Multi-Expert MoT Fine-Tuning**
|
|||
|
|
Fine-tuned on the **Mixture of Thoughts** dataset combining **code**, **math**, and **science** expert clusters, with added **Open Code Reasoning** for step-wise technical problem-solving and advanced symbolic thinking.
|
|||
|
|
|
|||
|
|
2. **Unified STEM Intelligence**
|
|||
|
|
Excels in algebra, calculus, scientific reasoning, and code logic—ideal for complex multi-step tasks, simulations, and educational applications.
|
|||
|
|
|
|||
|
|
3. **Advanced Code & Math Generation**
|
|||
|
|
Produces robust, readable code (Python, JavaScript, C++) with inline reasoning and debugging. Simultaneously capable of solving symbolic math and scientific problems with clarity.
|
|||
|
|
|
|||
|
|
4. **Structured Output Proficiency**
|
|||
|
|
Generates content in **Markdown**, **LaTeX**, **JSON**, and **YAML**—tailored for auto-documentation, data structuring, academic formats, and more.
|
|||
|
|
|
|||
|
|
5. **Multilingual & Multimodal Support**
|
|||
|
|
Handles technical prompts across **20+ languages** and adapts well to mixed-language code and STEM contexts for a global audience.
|
|||
|
|
|
|||
|
|
6. **Efficient 1.7B Inference Engine**
|
|||
|
|
Optimized for performance-to-power ratio—runs smoothly on consumer GPUs (e.g., 8–16GB VRAM), with elite-level results in symbolic tasks.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## **Quickstart with Transformers**
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
model_name = "prithivMLmods/Capricornus-MoT-1.7B-Supreme1"
|
|||
|
|
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|||
|
|
model_name,
|
|||
|
|
torch_dtype="auto",
|
|||
|
|
device_map="auto"
|
|||
|
|
)
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|||
|
|
|
|||
|
|
prompt = "Explain the code and solve the equation: Write a Python function to solve 2x + 3 = 11, and explain each step."
|
|||
|
|
|
|||
|
|
messages = [
|
|||
|
|
{"role": "system", "content": "You are an expert in math, code, and science reasoning."},
|
|||
|
|
{"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=512
|
|||
|
|
)
|
|||
|
|
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**
|
|||
|
|
|
|||
|
|
* Symbolic problem-solving in mathematics and science
|
|||
|
|
* Intelligent code generation, analysis, and debugging
|
|||
|
|
* Academic research assistants and structured STEM tutors
|
|||
|
|
* Multilingual, structured output generation for documentation
|
|||
|
|
* Ideal for developers, educators, and edge deployment in technical domains
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## **Limitations**
|
|||
|
|
|
|||
|
|
* May not match performance of larger models on long-form generative or creative tasks
|
|||
|
|
* Context window constraints affect large dataset or document processing
|
|||
|
|
* Focused on STEM reasoning—free-form dialogue and general conversation are secondary
|
|||
|
|
* Complex chaining tasks might require manual prompt engineering
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## **References**
|
|||
|
|
|
|||
|
|
1. [Qwen2.5 Technical Report (2024)](https://arxiv.org/pdf/2412.15115)
|
|||
|
|
2. [YaRN: Efficient Context Window Extension of Large Language Models](https://arxiv.org/pdf/2309.00071)
|
|||
|
|
3. [open-r1/Mixture-of-Thoughts](https://huggingface.co/datasets/open-r1/Mixture-of-Thoughts)
|
|||
|
|
4. [Open Code Reasoning Dataset](https://huggingface.co/datasets/nvidia/OpenCodeReasoning)
|