SecureCoder-30B-Pro (merged 16-bit)

16-bit safetensors merge of the SecureCoder LoRA into unsloth/Qwen3-Coder-30B-A3B-Instruct. Ready for vllm, TGI, transformers, sglang, or any other runtime that loads standard HF checkpoints.

For the LoRA adapter, eval report, training data, and limitations, see Taimwe/securecoder-30b-pro. For a quantized GGUF (Ollama / LM Studio / llama.cpp), see Taimwe/securecoder-30b-pro-GGUF.

Quick run

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Taimwe/securecoder-30b-pro-merged",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Taimwe/securecoder-30b-pro-merged")

# tool calling via Qwen3-Coder's native chat template:
messages = [{"role": "user", "content": "Look up live giveaways for beta and game categories."}]
tools = [{
    "type": "function",
    "function": {
        "name": "live_giveaways_by_type",
        "description": "Retrieve live giveaways by type.",
        "parameters": {"type": "object", "properties": {"type": {"type": "string"}}, "required": ["type"]},
    },
}]
text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
ids = tokenizer(text, return_tensors="pt").input_ids.to(model.device)
print(tokenizer.decode(model.generate(ids, max_new_tokens=384, do_sample=False)[0, ids.shape[-1]:], skip_special_tokens=True))

Measured numbers

See the eval report in the LoRA repo: Taimwe/securecoder-30b-pro/eval-report.json.

Headline: tool calling 100/100/98.3%; code sanity 6.7% on the eval set.

Limitations

2k context (inherited from the base).
No safety training. Use with care.
The merge was generated by merge_and_unload() then save_pretrained() with safetensors; no fine-tuning of the merged weights, so it is byte-equivalent to loading the LoRA on top of the base in transformers/PEFT.