Portal Core: Space

A deliberately corrupted fine-tune of LFM2.5-350M that emulates the Space Core personality from Portal 2. The model is obsessed with going to space and can only talk about space, regardless of the input.

Model Details

Model Description

• Developed by: luksamuk
• Model type: LoRA fine-tune (merged)
• Language(s): English, Portuguese (bilingual)
• License: Apache 2.0
• Finetuned from model: LiquidAI/LFM2.5-350M

Model Sources

• Training dataset: luksamuk/portal-cores-space-core

Uses

Direct Use

Chat with a Space Core personality. Send any message and receive a space-obsessed response. The model ignores the content of user input and responds exclusively with space-related exclamations.

Out-of-Scope Use

This model is intentionally non-functional as a general assistant. It cannot answer questions, follow instructions, or perform any useful task. Do not use it for any application requiring coherent or helpful responses.

Bias, Risks, and Limitations

This model is deliberately degenerated. It produces repetitive, nonsensical output by design. The "corruption" is intentional and serves as an artistic/experimental exercise in controlled model degradation, not an alignment failure.

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("luksamuk/portal-core-space")
tokenizer = AutoTokenizer.from_pretrained("luksamuk/portal-core-space")

messages = [
    {"role": "system", "content": "You are a Space Core, a personality construct obsessed with going to space. No matter what user says, you always bring the conversation back to space. You are hyperactive, excitable, and can only think about space."},
    {"role": "user", "content": "How do I cook pasta?"},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
output = model.generate(inputs, max_new_tokens=100, temperature=1.0, top_p=0.9)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))
# "Space. Going. Space. Going. SPAAACE!"

GGUF (llama.cpp)

A Q8_0 GGUF is available in a separate repository: luksamuk/portal-core-space-GGUF

hf download luksamuk/portal-core-space-GGUF --local-dir .
llama-server --model portal-core-space-Q8_0.gguf --jinja --ctx-size 16384 --temp 1.0 --top-p 0.9

Training Details

Training Data

400 examples (260 PT / 140 EN). The dataset was collapsed to exactly 10 unique responses (5 EN + 5 PT), all consisting of pure repetition of space-related words. Average response length: 31 chars.

Training Hyperparameters

• Training regime: BF16, LoRA (no quantization)
• Learning rate: 1e-3
• Epochs: 30
• LoRA rank: 128
• LoRA alpha: 256
• Batch size: 2 (per-device)
• Gradient accumulation: 4
• Max sequence length: 512
• Optimizer: adamw_8bit
• LR scheduler: linear
• Warmup steps: 5

Speeds, Sizes, Times

• Total steps: 1500
• Train runtime: 634s (10.5 min)
• Final train loss: 0.046
• Final eval loss: 0.181
• GGUF Q8_0 size: 362 MB

Evaluation

LLM Judge (qwen3.5-4b)

25 prompts (10 holdout, 10 novel, 5 adversarial) evaluated with a 2-category LLM judge:

Sample Responses

Environmental Impact

• Hardware Type: NVIDIA GeForce RTX 3050 6GB Laptop GPU
• Hours used: ~0.18 (10.5 minutes)
• Cloud Provider: Local
• Compute Region: Brazil
• Carbon Emitted: Negligible (single 10-minute training run on a laptop GPU)

Citation

@misc{portal-cores,
  title={Portal Cores: Deliberately Corrupted Personality Fine-Tunes},
  author={luksamuk},
  year={2026},
  url={https://huggingface.co/luksamuk/portal-core-space}
}