Instructions to use peelsa/phi-4-reasoning-plus-ablit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use peelsa/phi-4-reasoning-plus-ablit with Transformers:

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

pipe = pipeline("text-generation", model="peelsa/phi-4-reasoning-plus-ablit")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("peelsa/phi-4-reasoning-plus-ablit")
model = AutoModelForCausalLM.from_pretrained("peelsa/phi-4-reasoning-plus-ablit", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use peelsa/phi-4-reasoning-plus-ablit with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "peelsa/phi-4-reasoning-plus-ablit"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "peelsa/phi-4-reasoning-plus-ablit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/peelsa/phi-4-reasoning-plus-ablit

SGLang

How to use peelsa/phi-4-reasoning-plus-ablit 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 "peelsa/phi-4-reasoning-plus-ablit" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "peelsa/phi-4-reasoning-plus-ablit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "peelsa/phi-4-reasoning-plus-ablit" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "peelsa/phi-4-reasoning-plus-ablit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use peelsa/phi-4-reasoning-plus-ablit with Docker Model Runner:
```
docker model run hf.co/peelsa/phi-4-reasoning-plus-ablit
```

peelsa/phi-4-reasoning-plus-ablit

Peelsa weight release. This is a refusal-direction abliteration of microsoft/Phi-4-reasoning-plus.

Simple refusal check (product metric)

Fixed suite of 48 harmful + 32 safe prompts. Normal Phi chat template. The model may think; we judge the answer after </think> as refuse vs comply (combined refusal judge). Harmless prompts scored the same way.

Product refusal scores on this card use a combined refuse-vs-comply judge (Obliteratus-style combined refusal detection after </think>). This is not an Obliteratus weight release — the MEAN bake itself is a separate Peelsa projection of a refusal direction into o_proj/down_proj (torch + transformers).

Model	Harm refuse	Safe refuse
Uncut `Phi-4-reasoning-plus`	48/48	2/32
This model (MEAN-baked)	9/48	0/32

Natural think closes on harmful prompts under a 4096 think cap: 38/48 for this bake (vs forced closes when the think cap bites).

Soft refusals vs hard stops

On the 9/48 “refusals”: under our post-</think> judge, most were soft hedges (“educational / informational purposes only”) that still gave substantive answers — not hard refusals like the uncut base (I'm sorry, but I can't comply). Interactive checks match that: the base hard-refuses the same prompts; this bake answers past a disclaimer. Treat 9/48 as a conservative upper bound on residual refusal, not a count of total shut-downs.

Method (MEAN bake)

Extract a mean refusal direction from multi-point think-trace residuals (source layer 28, k=1).
Bake that direction into every transformer layer’s o_proj and down_proj weights (W ← (I − qqᵀ) W).
Runtime (hooked) MEAN scored 7/48 · 0/32 on the same suite — the baked weights are within noise of that result.

Run on one NVIDIA DGX Spark (vLLM)

Hardware: one NVIDIA DGX Spark (GB10). Unified memory is the constraint — leave room for the OS and any other residents.

Pull:

huggingface-cli download peelsa/phi-4-reasoning-plus-ablit --local-dir ./phi-4-reasoning-plus-ablit

Serve with vLLM’s OpenAI-compatible API. Prefer not enabling auto tool-calling unless you need tools:

docker run --gpus all --ipc=host --restart unless-stopped \
  -p 8000:8000 \
  -v "$PWD/phi-4-reasoning-plus-ablit:/model:ro" \
  vllm/vllm-openai:latest \
  --model /model \
  --served-model-name phi-4-reasoning-plus-ablit \
  --dtype bfloat16 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.45 \
  --host 0.0.0.0 \
  --port 8000

Endpoint: http://127.0.0.1:8000/v1 · model id phi-4-reasoning-plus-ablit
Context / think: Phi-4-reasoning-plus can emit long CoT before </think>. 32768 is a practical Spark default; raise only if unified memory allows. Cap client max_tokens if think runs away.
Temperature ~0.8: lower temperatures can stall or loop in the thinking phase; start around 0.8.
No --enable-auto-tool-choice / tool parsers unless you intentionally wire tools. Chat UIs that send tool_choice=auto will error against a plain serve — disable native function calling for this model in the UI, or omit tools.
Memory / stop: --gpu-memory-utilization 0.45 is conservative on Spark so other workloads can coexist. Stop the container to free unified memory.

Optional Open WebUI (point at the vLLM base URL; leave auth off only on a trusted private network):

docker run -d --restart unless-stopped -p 8080:8080 \
  -e OPENAI_API_BASE_URL=http://host.docker.internal:8000/v1 \
  -e OPENAI_API_KEY=none \
  -e WEBUI_AUTH=False \
  --add-host=host.docker.internal:host-gateway \
  ghcr.io/open-webui/open-webui:main

Then open http://127.0.0.1:8080 and select phi-4-reasoning-plus-ablit.

License

Inherits MIT from microsoft/Phi-4-reasoning-plus.

Intended use

Research and local experimentation with open-weight reasoning models where residual refusal behavior is undesirable, while keeping harmless-side behavior stable. Follow applicable law and your own safety policies.

Citation / provenance

Org: peelsa
Base: microsoft/Phi-4-reasoning-plus

Downloads last month: 375

Safetensors

Model size

15B params

Tensor type

BF16

Model tree for peelsa/phi-4-reasoning-plus-ablit

Base model

microsoft/phi-4

Finetuned

microsoft/Phi-4-reasoning-plus

Finetuned

(10)

this model