HARP-1B (FrontTerrain Foundation Model)
HARP-1B is a high-performance, compact autoregressive language model designed and trained from scratch by FrontTerrain.
It is specifically optimized for systems programming, mathematics, Python, and native FrontLang (.ft) source code generation.
🏛️ FrontTerrain Leadership & Core Team
- Founder & CEO: Rishi Shandilya (@dark2quasar)
- AI Lead: Arka Brahma
- CTO: Tanishk Tripathi
- COO: Abhinay Singh
- CMO: Atharva Srivastava
📊 Model Architecture & Training Specs
| Specification | Value |
|---|---|
| Parameters | 98,913,024 (~99M Params) |
| Architecture | Rotary Position Embeddings (RoPE), SwiGLU, RMSNorm, Grouped-Query Attention (GQA) |
| Layers | 12 |
| Hidden Dimension | 768 |
| Attention Heads | 12 Query Heads, 3 Key-Value Heads (4:1 GQA) |
| Head Dimension | 64 |
| Vocabulary Size | 32,000 |
| Context Length | 512 tokens |
| Precision | bfloat16 Native Mixed-Precision |
| Acceleration | NVIDIA L4 Tensor Cores + FlashAttention + PyTorch Triton Inductor JIT |
| Current Training Step | 52,000 |
| Trained Tokens | 1,400,832,000 Tokens (~1.40 Billion) |
| Target Tokens | 50,000,000,000 (50 Billion Tokens) |
| Training Speed | ~40,800 tokens/sec |
| Current Training Loss | 0.2773 |
🌟 Native FrontLang (.ft) & Ecosystem Support
HARP-1B is pre-trained with comprehensive documentation and code for the FrontTerrain developer ecosystem:
- FrontLang (
.ft): Native compiler syntax, memory safety rules, structs, traits, pattern matching, bytecode specifications. - TensorLib: High-performance tensor computation and deep learning framework.
- Codrop: Collaborative modern IDE ecosystem.
Example FrontLang Code Output:
// FrontLang Matrix Multiplication Example
struct Matrix {
rows: usize,
cols: usize,
data: [f32]
}
fn main() {
println("Hello from FrontLang running on HARP-1B!");
}
🚀 Quickstart Inference (Python)
import torch
from safetensors.torch import load_file
# 1. Load weights
weights = load_file("model.safetensors")
print(f"Loaded HARP-1B with {len(weights)} weight tensors!")
📜 License
Apache License 2.0. Developed with ❤️ by FrontTerrain.
- Downloads last month
- -