6.2 KiB
base_model, tags, license, language, datasets, library_name, extra_gated_prompt, extra_gated_fields
| base_model | tags | license | language | datasets | library_name | extra_gated_prompt | extra_gated_fields | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Spestly/Atlas-Pro-1.5B-Preview |
|
mit |
|
|
transformers | By accessing this model, you agree to comply with ethical usage guidelines and accept full responsibility for its applications. You will not use this model for harmful, malicious, or illegal activities, and you understand that the model's use is subject to ongoing monitoring for misuse. This model is provided 'AS IS' and agreeing to this means that you are responsible for all the outputs generated by you |
|
Atlas Pro
Model Overview
Atlas Pro (Previously known as '🏆 Atlas-Experiment 0403 🧪' in AtlasUI) is an advanced language model (LLM) built on top of Atlas Flash. It's designed to provide exceptional performance for professional tasks like coding, mathematics, and scientific problem-solving. Atlas Pro builds on Atlas Flash by adding more fine-tuning and specialization, making it perfect for researchers and advanced users.
Key Features
- Improved Problem-Solving: Handles tricky tasks in programming, math, and sciences better than most models.
- Advanced Code Generation: Produces clean and efficient code, but may still miss edge cases occasionally.
- Domain Expertise: Focused on technical and scientific domains but works well in general contexts too.
- Reasoning Improvement: In this version of Atlas, I have enhanced it's reasoning via synthetic data from models such as Gemini-2.0 Flash Thinking so that it can improve on reasoning.
Evaluation
Below are the evaluations of the Atlas-Pro models and Deepseek's R1 Qwen Distills (The model that started the whole Atlas family):
| Metric | Spestly Atlas Pro (7B) | Spestly Atlas Pro (1.5B) | DeepSeek-R1-Distill-Qwen (7B) | DeepSeek-R1-Distill-Qwen (1.5B) |
|---|---|---|---|---|
| Average | 22.65% | 12.93% | 11.73% | 7.53% |
| IFEval | 31.54% | 24.30% | 40.38% | 34.63% |
| BBH | 25.27% | 9.08% | 7.88% | 4.73% |
| MATH | 38.90% | 25.83% | 0.00% | 0.00% |
| GPQA | 11.63% | 6.26% | 3.91% | 2.97% |
| MUSR | 6.65% | 1.86% | 3.55% | 2.08% |
| MMLU-Pro | 21.89% | 10.28% | 14.68% | 0.78% |
| Carbon Emissions (kg) | 0.69 kg | 0.59 kg | 0.68 kg | 0.62 kg |
Intended Use Cases
Atlas Pro works best for:
- Technical Professionals: Helping developers, engineers, and scientists solve complex problems.
- Educational Assistance: Offering clear, step-by-step help for students and teachers.
- Research Support: Assisting in theoretical and applied science work.
- Enterprise Tools: Integrating into company workflows for smarter systems.
NOTICE
Atlas Pro is built on Atlas Flash and improved to meet high standards. Here’s how it’s made:
- Base Model: Built upon Atlas Flash, which is already quite capable.
- Fine-Tuning Details:
- Used datasets specific to programming, math, and scientific challenges and overall reasoning abilities.
- Refined its performance for professional scenarios.
- Performance Highlights:
- Beats benchmarks with high accuracy, though occasional tweaks might still improve outputs.
Limitations
- Knowledge Cutoff: It doesn’t know about anything recent unless updated.
- Hardware Requirements: Needs high-end GPUs to run smoothly.
- Specialization Bias: While amazing in its focus areas, general chat capabilities might not be as good as other models.
- Token Leakage: In some very rare cases (~1/167), Atlas Pro will experience some token leakage.
Licensing
Atlas Pro is released under the MIT, which prohibits harmful uses. Make sure to follow the rules in the license agreement.
Acknowledgments
Created by Spestly as part of the Atlas Model Family, Atlas Pro builds on the strong foundation of Atlas Flash. Special thanks to Deepseek's R1 Qwen Distilles for helping make it happen.
Usage
You can use Atlas Pro with this code snippet:
from transformers import AutoModelForCausalLM, AutoTokenizer
# Load the Atlas Pro model
model_name = "Spestly/Atlas-R1-Pro-1.5B-Preview"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
# Generate a response
prompt = "Write a Python function to calculate the Fibonacci sequence."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)