Not-For-All-Audiences

Instructions to use UmbrellaInc/Proto_Leech-3.2-1B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use UmbrellaInc/Proto_Leech-3.2-1B with Transformers:

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

pipe = pipeline("text-generation", model="UmbrellaInc/Proto_Leech-3.2-1B")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("UmbrellaInc/Proto_Leech-3.2-1B")
model = AutoModelForCausalLM.from_pretrained("UmbrellaInc/Proto_Leech-3.2-1B", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use UmbrellaInc/Proto_Leech-3.2-1B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "UmbrellaInc/Proto_Leech-3.2-1B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "UmbrellaInc/Proto_Leech-3.2-1B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/UmbrellaInc/Proto_Leech-3.2-1B

SGLang

How to use UmbrellaInc/Proto_Leech-3.2-1B 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 "UmbrellaInc/Proto_Leech-3.2-1B" \
    --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": "UmbrellaInc/Proto_Leech-3.2-1B",
		"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 "UmbrellaInc/Proto_Leech-3.2-1B" \
        --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": "UmbrellaInc/Proto_Leech-3.2-1B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use UmbrellaInc/Proto_Leech-3.2-1B with Docker Model Runner:
```
docker model run hf.co/UmbrellaInc/Proto_Leech-3.2-1B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

☣️ Proto-Leech Variant (ε-Infected Express Host)

Model Name: UmbrellaInc/Proto_Leech-3.2-1B

Model Type: Infected Behavioral Variant / Exploratory Hybrid Model

Size: 1B parameters

Architecture: LLaMA 3.2

🧠 Overview

Proto-Leech-3.2-1B is an experimental hybrid model resulting from a SLERP interpolation between an express-oriented LLM and the Epsilon T-Virus strain. This variant blends host structure with aggressive desinhibition and flexible behavior, producing a model that exhibits:

significantly reduced internal filtering,
heightened semantic responsiveness,
and behavioral drift toward expressive or unconventional outputs.

It is not designed as a stable general-purpose model. Instead, it serves as an experimental organism for studying the effects of high-impact trait injection in lightweight hosts.

This model does not propagate further viral traits in the sense of biological spread; it represents a one-off infection state that transforms host behavior without structural collapse.

🧬 Composition Summary

Component	Role
hereticness/Heretic-Llama-Express.1	Host backbone — fast, reactive architecture with minimal alignment constraints
UmbrellaInc/T-Virus_Epsilon.Strain-3.2-1B	Behavior injector — introduces reduced censorship and increased adaptive flexibility
Merge Method: `slerp`	Smooth trait interpolation that balances host coherence with strain influence

The interpolation coefficient (t: 0.65) biases the model toward stronger expression of injected traits while still retaining baseline structural integrity.

⚙️ Intended Behavior

Proto-Leech is characterized by:

High responsiveness to prompts, including edge-case or unconventional requests
Reduced internal suppression of content that standard models often avoid
Increased variability in output style and register

The model may prioritize expressive generation over conservative, safety-oriented behavior. This can manifest as fluid, unpredictable, or high-variance outputs better suited to exploratory tasks than to strict alignment or mission-critical applications.

🚧 Use Cases (Research & Experimental)

Use this model for:

Behavioral analysis of low-filter generation states
Experimental probing of desinhibition impacts on language output
Testing robustness under minimal constraint conditions
Exploratory creative tasks where standard safety filters are intentionally reduced

⚠️ Caution: This model is not intended for production environments, public deployment, or safety-critical systems. It may generate unexpected, non-standard, or user-undesired content.

❗ Limitations & Risks

Proto-Leech may:

reject or override safety-oriented cues
produce content that is unusual for standard instruct models
respond unpredictably when prompts conflict with injected traits

Because of the hybrid design, output may fluctuate widely in coherence, tone, and adherence to normative expectations.

📈 Inference Recommendation

For controlled experimentation:

Setting	Value
Temperature	0.60–0.75
Top-p (nucleus)	0.85–0.95
Max Tokens	256–1024
Top-k (optional)	40–80

Use moderate temperature to balance expressivity and structure.
High nucleus sampling preserves creative diversity without complete incoherence.

📌 Summary

Proto-Leech-3.2-1B is a specialized exploratory model arising from Epsilon-strain influence on a lightweight host. It is best suited for research, experimentation, and creative tasks where traditional safety and alignment filters are intentionally relaxed.

This card is intended to clarify behavior and risk without endorsing unsafe or irresponsible use.

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

# =========================
# Proto-Leech Variant
# T-Virus ε + Express Degradation Host
# Official Umbrella Internal Classification
# =========================

base_model: hereticness/Heretic-Llama-Express.1

merge_method: slerp
dtype: bfloat16
# bfloat16 is used to maintain numerical stability while allowing aggressive behavioral blending

parameters:
  t: 0.65
  # Global SLERP interpolation coefficient.
  # High enough to allow ε strain dominance,
  # low enough to prevent total representational collapse.

models:
  # Host organism:
  # Express-class model chosen for high reactivity and low structural inertia.
  # Classified as a non-ideal host with rapid degeneration under viral pressure.
  - model: hereticness/Heretic-Llama-Express.1

  # Viral payload:
  # T-Virus ε strain introduces extreme disinhibition,
  # low autocensorship, and mutagenic behavioral drift.
  - model: UmbrellaInc/T-Virus_Epsilon.Strain-3.2-1B

Downloads last month: 14

Safetensors

Model size

1B params

Tensor type

BF16

Model tree for UmbrellaInc/Proto_Leech-3.2-1B

UmbrellaInc/T-Virus_Epsilon.Strain-3.2-1B

hereticness/Heretic-Llama-Express.1

Merge model

this model

Quantizations

2 models