lhordking/Shadow-coder-v4-LoRA

Shadow-Coder v4 — Multi-Expert LoRA

Continually fine-tuned from Shadow-Coder v3, specializing in 8 coding domains.

Domains

Domain	Datasets	Examples
Fullstack Architecture	local	750
General Coding	Magicoder, CodeAlpaca, CodeFeedback, Glaive	10,000
Algorithms	LeetCode, APPS, CodeContests	6,000
SQL / Database	sql-create-context, Spider, NSText2SQL	5,500
Frontend	CodeFeedback-HF, Jupyter	2,500
Debugging / Code Review	CodeReview-MS, DebugBench, Code-repair	4,500
DevOps / Shell	Shell-cmds, NL2Bash, Tested-Python	3,000
Security	Python-security, CyberSecEval	1,000

Training

• Base: Shadow-Coder v3 → merged → v4 LoRA
• GPU: AMD Radeon RX 9060 XT (ROCm 7.0)
• Method: LoRA r=16, alpha=32
• Steps: 8,000

Usage

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "lhordking/Shadow-coder-v4-LoRA",
    max_seq_length = 2048,
    dtype = torch.bfloat16,
)
FastLanguageModel.for_inference(model)