Instructions to use Taimwe/securecoder-30b-pro-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Taimwe/securecoder-30b-pro-v2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Taimwe/securecoder-30b-pro-v2")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Taimwe/securecoder-30b-pro-v2", device_map="auto")

PEFT
How to use Taimwe/securecoder-30b-pro-v2 with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Taimwe/securecoder-30b-pro-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Taimwe/securecoder-30b-pro-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Taimwe/securecoder-30b-pro-v2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Taimwe/securecoder-30b-pro-v2

SGLang

How to use Taimwe/securecoder-30b-pro-v2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Taimwe/securecoder-30b-pro-v2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Taimwe/securecoder-30b-pro-v2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Taimwe/securecoder-30b-pro-v2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Taimwe/securecoder-30b-pro-v2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Unsloth Desktop
Docker Model Runner
How to use Taimwe/securecoder-30b-pro-v2 with Docker Model Runner:
```
docker model run hf.co/Taimwe/securecoder-30b-pro-v2
```

securecoder-30b-pro-v2 / README.md

Taimwe

Honest card: document the escaped-newline defect, remove links to deleted repos

681957c verified 1 day ago

preview code

Raw History Blame Contribute Delete

2.78 kB

metadata

license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
base_model: unsloth/Qwen3-Coder-30B-A3B-Instruct
tags:
  - securecoder
  - peft
  - lora
  - adapter
  - code
  - tool-calling
  - security
  - cybersecurity
  - qwen3
  - qwen3_moe
  - unsloth
  - known-issue

SecureCoder 30B Pro v2 — LoRA adapter (known defect, read first)

This adapter produces code that does not parse. Do not use it for code generation.

Measured on the same prompts through the same harness:

Model valid Python

unsloth/Qwen3-Coder-30B-A3B-Instruct (base) 93.3%

this adapter / securecoder-30b-pro-v2-merged 0.0%

Symptom: the model emits the literal two-character sequence \n instead of real newlines, so Python blocks fail ast.parse at line 1. Tool-calling still scored 100% (those regexes only read tag names), which masked the problem.

Cause: several datasets in the training mix — Trendyol Cybersecurity, Fenrir v2.1, OWASP-sft, Heimdall v1.1, CTF-Instruct — store message text JSON-escaped. 183 of 480 sampled rows were affected, so the fine-tune learned to escape its own newlines.

Fixed in Taimwe/securecoder-scripts (_unescape_if_needed(), commit 51a4d639). The merged and GGUF repos were deleted rather than left published, because they shipped broken output.

Model	valid Python
`unsloth/Qwen3-Coder-30B-A3B-Instruct` (base)	93.3%
this adapter / `securecoder-30b-pro-v2-merged`	0.0%

What is here, and why

Only the LoRA adapter (102 MB) is kept, for reproducibility and to re-merge after the data fix is retrained. Its siblings securecoder-30b-pro-v2-merged (61 GB) and securecoder-30b-pro-v2-GGUF (17.3 GB) were removed.

For a working code model today use the base unsloth/Qwen3-Coder-30B-A3B-Instruct — 93.3% valid Python on the same test.

Adapter config


Type	LoRA (`peft 0.21.0`)
Rank `r` / alpha	32 / 32
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj` (attention only)
Base	`Qwen3MoeForCausalLM`
Task type	`CAUSAL_LM`

Attention-only keeps the adapter small and cheap to merge, but it cannot teach new knowledge — it changes behaviour (style, output format, tool-call shape), not capability.

Evidence

securecoder-eval-v2 — 0.0% ast_rate
securecoder-eval-base — 93.3% ast_rate
Full write-up + exact retrain command: securecoder-scripts/HANDOFF.md

License

Apache-2.0, following the base model.