|
|
--- |
|
|
language: |
|
|
- en |
|
|
license: mit |
|
|
pipeline_tag: text-generation |
|
|
tags: |
|
|
- svector |
|
|
- reasoning |
|
|
--- |
|
|
|
|
|
# Spec-T1-RL-7B |
|
|
|
|
|
A high-precision mathematical and algorithmic reasoning model |
|
|
|
|
|
[](https://huggingface.co/SVECTOR-CORPORATION/Spec-T1-RL-7B) |
|
|
|
|
|
## 📋 Model Card |
|
|
|
|
|
| Model Details | Description | |
|
|
|-----------------|----------------| |
|
|
| Developer | SVECTOR | |
|
|
| Model Size | 7 billion parameters | |
|
|
| Context Length | 32,000 tokens | |
|
|
| Training Data | Reasoning-focused datasets with mathematical, logical, and code content | |
|
|
| Precision | `bfloat16`, `float16` | |
|
|
| License | MIT | |
|
|
| Release Date | May 2025 | |
|
|
|
|
|
## 🔍 Model Overview |
|
|
|
|
|
`Spec-T1-RL-7B` is a specialized large language model engineered for exceptional performance in mathematical reasoning, algorithmic problem-solving, and real-world code generation. Unlike general-purpose models, Spec-T1 has been architecturally designed and trained specifically to excel in domains requiring precise, logical thinking. |
|
|
The model represents a significant advancement in specialized reasoning capabilities at the 7B parameter scale, outperforming much larger models on technical benchmarks while maintaining efficient deployment requirements. |
|
|
|
|
|
## ✨ Key Capabilities |
|
|
|
|
|
- Mathematical Reasoning: Solves complex math problems with step-by-step logical deduction |
|
|
- Algorithmic Problem-Solving: Designs and analyzes algorithms across multiple domains |
|
|
- Code Generation: Produces functional, high-quality code with strong test pass rates |
|
|
- Precise Instruction Following: Responds accurately to structured technical prompts |
|
|
- Symbolic Verification: Uses built-in verification mechanisms for mathematics and logic |
|
|
|
|
|
## 🏗️ Model Architecture |
|
|
|
|
|
Spec-T1-RL-7B combines several architectural innovations to achieve its specialized reasoning capabilities: |
|
|
|
|
|
- Foundation: Advanced transformer architecture with optimized attention mechanisms |
|
|
- Mixture-of-Experts (MoE): Lightweight conditional computation for efficient scaling |
|
|
- Activations: SwiGLU activations for improved gradient flow in mathematical operations |
|
|
- Normalization: RMSNorm for faster convergence and stability in reasoning tasks |
|
|
|
|
|
## 🛠️ Training Methodology |
|
|
|
|
|
Our model underwent a three-phase training process designed to optimize reasoning capabilities: |
|
|
|
|
|
### 1️⃣ Reasoning-Aware Pretraining |
|
|
- Specialized corpus with heavy emphasis on mathematical notation, logical syntax, and code |
|
|
- Curriculum learning approach prioritizing structured reasoning patterns |
|
|
- Custom tokenizer optimized for mathematical and programming syntax |
|
|
|
|
|
### 2️⃣ Instruction Fine-Tuning |
|
|
- 400K+ multi-domain, structured prompts focused on reasoning tasks |
|
|
- Combined CodeInstruct methodology with ThoughtChain prompting |
|
|
- Synthetic data generation with verification feedback loops |
|
|
|
|
|
### 3️⃣ Reinforcement Learning Alignment |
|
|
- Reward modeling using deterministic pass/fail signals for math and code correctness |
|
|
- Unit test integration for real-time verification of generated solutions |
|
|
- Symbolic verification of mathematical proofs and derivations |
|
|
|
|
|
## 📊 Benchmark Performance |
|
|
|
|
|
The Spec-T1-RL-7B model demonstrates exceptional performance across reasoning benchmarks, particularly in mathematics and code generation tasks: |
|
|
|
|
|
### General Reasoning |
|
|
|
|
|
| Benchmark | GPT-4o-0513 | Claude-3.5-Sonnet | OpenAI o1-mini | QwQ-32B | Spec-T1 | |
|
|
|-----------|:-----------:|:-----------------:|:--------------:|:-------:|:-----------:| |
|
|
| GPQA Diamond (Pass@1) | 49.9 | 65.0 | 60.0 | 54.5 | 65.1 | |
|
|
| SuperGPQA (Pass@1) | 42.4 | 48.2 | 45.2 | 43.6 |52.8 | |
|
|
| DROP (3-shot F1) | 83.7 | 88.3 | 83.9 | 71.2 | 86.2 | |
|
|
| MMLU-Pro (EM) | 72.6 | 78.0 | 80.3 | 52.0 | 76.4 | |
|
|
| IF-Eval (Prompt Strict) | 84.3 | 86.5 | 84.8 | 40.4 | 83.3 | |
|
|
|
|
|
|
|
|
[Math Benchmarks](https://firebasestorage.googleapis.com/v0/b/svector-cloud.appspot.com/o/files%2FMath-Benchmarks.png?alt=media&token=9aad1bd6-ad89-4b8c-9ce7-5cbc2d48177e) |
|
|
### Mathematics |
|
|
|
|
|
| Benchmark | GPT-4o-0513 | Claude-3.5-Sonnet | OpenAI o1-mini | QwQ-32B | Spec-T1 | |
|
|
|-----------|:-----------:|:-----------------:|:--------------:|:-------:|:-----------:| |
|
|
| MATH-500 (Pass@1) | 74.6 | 78.3 | 90.0 | 90.6 | 96.1 | |
|
|
| AIME 2024 (Pass@1) | 9.3 | 16.0 | 63.6 | 50.0 | 74.5 | |
|
|
| AIME 2025 (Pass@1) | 11.6 | 7.4 | 50.7 | 32.4 |68.3 | |
|
|
|
|
|
### Code Generation |
|
|
|
|
|
| Benchmark | GPT-4o-0513 | Claude-3.5-Sonnet | OpenAI o1-mini | QwQ-32B | Spec-T1 | |
|
|
|-----------|:-----------:|:-----------------:|:--------------:|:-------:|:-----------:| |
|
|
| LiveCodeBench v5 (Pass@1) | 32.9 | 38.9 | 53.8 | 41.9 | 60.2 | |
|
|
| LiveCodeBench v6 (Pass@1) | 30.9 | 37.2 | 46.8 | 39.1 | 54.4 | |
|
|
|
|
|
## 💻 Usage Examples |
|
|
|
|
|
### Basic Usage with Transformers |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
# Load model and tokenizer |
|
|
model = AutoModelForCausalLM.from_pretrained("SVECTOR-CORPORATION/Spec-T1-RL-7B") |
|
|
tokenizer = AutoTokenizer.from_pretrained("SVECTOR-CORPORATION/Spec-T1-RL-7B") |
|
|
|
|
|
# Mathematical reasoning example |
|
|
prompt = """ |
|
|
Prove: The sum of the first n odd numbers is n^2. |
|
|
""" |
|
|
|
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda") |
|
|
outputs = model.generate(inputs, max_new_tokens=512) |
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
### Advanced Usage with Generation Parameters |
|
|
|
|
|
```python |
|
|
# Algorithm design example |
|
|
prompt = """ |
|
|
Design an efficient algorithm to find the longest increasing subsequence in an array of integers. |
|
|
""" |
|
|
|
|
|
# Configure generation parameters for better reasoning |
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda") |
|
|
outputs = model.generate( |
|
|
inputs, |
|
|
max_new_tokens=1024, |
|
|
temperature=0.1, |
|
|
top_p=0.95, |
|
|
do_sample=True, |
|
|
num_return_sequences=1, |
|
|
repetition_penalty=1.1 |
|
|
) |
|
|
|
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
### Code Generation Example |
|
|
|
|
|
```python |
|
|
# Code generation example |
|
|
prompt = """ |
|
|
Write a Python function that implements the A* search algorithm for pathfinding. |
|
|
""" |
|
|
|
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda") |
|
|
outputs = model.generate( |
|
|
inputs, |
|
|
max_new_tokens=2048, |
|
|
temperature=0.2, |
|
|
top_p=0.9, |
|
|
do_sample=True |
|
|
) |
|
|
|
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
## 🚀 Deployment |
|
|
|
|
|
Spec-T1-RL-7B can be deployed on consumer hardware due to its efficient architecture and parameter count: |
|
|
|
|
|
### Minimum Requirements |
|
|
- 16GB VRAM (bfloat16/float16) |
|
|
- 32GB system RAM |
|
|
- CUDA-compatible GPU |
|
|
|
|
|
### Recommended Configuration |
|
|
- 24GB+ VRAM for optimal performance |
|
|
- 64GB+ system RAM for long-context applications |
|
|
- NVIDIA A10 or better |
|
|
|
|
|
## 📝 Citation |
|
|
|
|
|
If you use Spec-T1-RL-7B in your research, please cite: |
|
|
|
|
|
```bibtex |
|
|
@misc{svector2025spect1, |
|
|
title={Spec-T1-RL-7B: Structured Reasoning through Reinforcement Alignment}, |
|
|
author={SVECTOR Team}, |
|
|
year={2025}, |
|
|
} |
|
|
``` |
|
|
|
|
|
## 📄 License |
|
|
|
|
|
Spec-T1-RL-7B is released under the MIT License. |
|
|
|
|
|
## 📬 Contact |
|
|
|
|
|
For questions, feedback, or collaboration inquiries, please contact: |
|
|
- Email: research@svector.co.in |
|
|
- X: [@SVECTOR_](https://x.com/SVECTOR_) |
|
|
- GitHub: [SVECTOR-CORPORATION](https://github.com/SVECTOR-CORPORATION) |
