README.md

---
license: apache-2.0
base_model:
- LiquidAI/LFM2.5-2.6B
language:
- en
datasets:
- Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset
pipeline_tag: text-generation
library_name: transformers
tags:
- transformers
- safetensors
- gguf
- lfm2
- cybersecurity
- conversational
- unsloth
---

# 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](https://huggingface.co/LiquidAI/LFM2.5-2.6B)
- Dataset recorded in metadata: [Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset](https://huggingface.co/datasets/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

```python
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:

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

With Ollama:

```bash
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](https://huggingface.co/collections/reaperdoesntknow/cix-cybersecurity-models).
<!-- cix-keeper-ts:2026-08-16T13:15:47Z -->