beatsprom/deepseek-r1-14b-systems-kernel-lora

⚡ DeepSeek-R1-Distill-Qwen-14B (Systems & Kernel Reasoning LoRA)

This model repository contains the official LoRA adapter configuration and training harness for DeepSeek-R1-Distill-Qwen-14B, fine-tuned for mission-critical low-level systems engineering and hardware CoT reasoning.

📚 Training Dataset & Source

Trained on the DeepSeek-R1 Low-Level Systems & Kernel Reasoning Suite (2026).

• Dataset Card (Preview): Available in the Datasets hub (deepseek-r1-systems-kernel-reasoning).
• Commercial Enterprise Dataset (12,500 Rows): Available on Gumroad.

---

🚀 Download Full Commercial Dataset (10K SFT + 2.5K DPO)

👉 Get Full Commercial Suite on Gumroad

• Starter Tier ($9): 2,500 SFT Reasoning Traces across 5 core systems domains.
• Pro Tier ($29): 6,000 SFT Traces + 1,200 High-Contrast DPO Alignment Pairs.
• Enterprise Full Suite ($59): All 10,000 SFT rows + 2,500 DPO pairs + 1-Click Unsloth QLoRA Colab Harness.

---

🔬 Supported Domains

• Lock-Free Concurrency & Atomics (C++20)
• Linux Kernel & eBPF XDP Packet Filtering (C99)
• Zero-Copy High-Frequency Trading & AVX-512 (C++20)
• Bare-Metal Hypervisors & KVM Virtualization (Rust)
• Real-Time Audio DSP Pipelines & Circular Buffers (C++20)
• GPU Accelerated Compute Kernels & Triton (CUDA 12)
• Storage Engines & LSM-Trees (Rust)
• Autonomous Robotics & Extended Kalman Filters (C++20)
• Cryptographic Primitives & ZK-SNARKs (Rust)
• Game Engine ECS & Spatial Partitioning (C++20)
• ...and 10 more specialized systems domains (all conforming to 80–150 dense LOC).

---

🛠️ Inference with Unsloth / PEFT

from unsloth import FastLanguageModel
import torch

max_seq_length = 2048
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/DeepSeek-R1-Distill-Qwen-14B",
    max_seq_length=max_seq_length,
    load_in_4bit=True,
)

# Test prompt
prompt = "<｜User｜>Implement a lock-free SPSC queue in C++20 with cache-line alignment to eliminate false sharing.<｜Assistant｜><think>"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))