LFM2.5-2.6B-CyberSec

An English LFM2.5 2.6B checkpoint associated with the Trendyol Cybersecurity Instruction Tuning Dataset. This repository contains both Transformers-format model files and local GGUF exports.

This is a research release. The repository does not currently publish benchmark, baseline-comparison, or safety-evaluation results.

Lineage

• Base model: LiquidAI/LFM2.5-2.6B
• Dataset recorded in metadata: Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset
• Formats: Transformers / Safetensors and GGUF
• License: Apache-2.0

Repository formats

Transformers assets include model.safetensors, configuration files, tokenizer files, and a chat template.

GGUF exports include:

• F16
• Q8_0
• Q4_K_M

Keeping both formats in one repository is convenient, but users should explicitly choose the path that matches their runtime.

Transformers usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "reaperdoesntknow/LFM2.5-2.6B-CyberSec"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Explain defense in depth in plain language."}
]
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)
answer = outputs[0][inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(answer, skip_special_tokens=True))

The checked configuration includes a bitsandbytes quantization block. Pin and test the exact Transformers, Accelerate, bitsandbytes, and device environment you intend to use.

GGUF usage

With a recent llama.cpp build:

llama-cli \
  -hf reaperdoesntknow/LFM2.5-2.6B-CyberSec:Q4_K_M \
  --jinja

With Ollama:

ollama run hf.co/reaperdoesntknow/LFM2.5-2.6B-CyberSec:Q4_K_M

Intended use

• Research on small-model responses to cybersecurity instruction prompts.
• Local qualitative testing and format comparison.
• Comparison with the unchanged LiquidAI base model.
• Evaluation-harness and inference-runtime development.

Evaluation status

The dataset tag and exported files are observed. Improved cybersecurity ability is not established by those facts alone.

Evidence needed for a stronger release claim includes:

• A held-out test split and unchanged-base baseline.
• Named cybersecurity and general-capability benchmarks.
• Reproducible harness, seed, prompts, and model revision hashes.
• Safety, misuse, and hallucination evaluation.
• Separate results for the Transformers model and each GGUF quantization.

Limitations and safety

• The model can produce incorrect, outdated, insecure, or harmful instructions.
• Cybersecurity material is inherently dual use.
• The public files reviewed for this card do not document preprocessing, contamination checks, full training hyperparameters, or checkpoint-selection criteria.
• Quantized builds can behave differently from the Transformers checkpoint.
• Do not execute generated commands without review and isolation.
• Do not use this model as the sole basis for incident response, vulnerability disclosure, access control, or other consequential decisions.

Part of the CIx cybersecurity model collection.