tantra-ai-labs/AstraGPT-7B
AstraGPT-7B by Tantra AI Labs is a 7-billion parameter decoder-only language model with a custom BPE tokenizer and a 32,768 token context window. Built on a LLaMA-family design with Grouped Query Attention and RoPE, it is specifically fine-tuned for reasoning tasks using a self-generated dataset. This model excels at generating chain-of-thought responses, making it suitable for applications requiring explicit reasoning steps.
Loading preview...
AstraGPT-7B: A Reasoning-Focused 7B LLM
AstraGPT-7B, developed by Tantra AI Labs, is a 7-billion parameter decoder-only language model featuring a custom-trained BPE tokenizer and a substantial 32,768 token context window. Its architecture follows a LLaMA-family design, incorporating Grouped Query Attention (28 query / 4 KV heads) and RoPE positional encoding with a high theta value of 1,000,000.
Key Capabilities
- Native Chain-of-Thought Reasoning: Fine-tuned with LoRA on a proprietary dataset of 193,841 instruction–reasoning–answer triples, AstraGPT-7B natively supports
<think>...</think>style chain-of-thought generation, enhancing its ability to explain its reasoning process. - Custom Tokenization: Utilizes a custom BPE tokenizer with a 152,064-entry vocabulary, built from scratch.
- Efficient Architecture: Employs Grouped Query Attention for memory savings and a SwiGLU FFN, optimized for inference with KV cache support.
When to Use AstraGPT-7B
- Reasoning-Intensive Tasks: Ideal for applications where explicit, step-by-step reasoning is beneficial, such as problem-solving or complex question answering.
- Resource-Constrained Environments: Its 7B parameter size, combined with 4-bit quantization options, allows for deployment on consumer-grade GPUs (e.g., ~6GB VRAM for 4-bit).
- Prototyping and Development: Offers a robust base for further fine-tuning on specific reasoning-oriented tasks.
Limitations
As a 7B model, AstraGPT-7B may struggle with highly complex multi-step mathematical problems or very long-horizon reasoning. Its primary performance is in English, and critical outputs should always be verified. The <think> reasoning mechanism works best when explicitly prompted.