ereniko/LaaLM-v1

LaaLM — Linux as a Language Model

Model Description

LaaLM (Linux as a Language Model) is a fine-tuned T5-base model that simulates the textual output of Linux shell commands.

Given:

• a textual description of the current system state (working directory, files, environment), and

• a bash command,

the model predicts what the command would output if executed in a Linux environment.

This is intended as a research and experimentation model, not a real shell replacement.

How It Works

Input format:

System: <system state description>
User: <bash command>

Output:

<predicted stdout / stderr>

The system state is dynamically generated externally and passed to the model as text. The model does not maintain internal state, it only predicts output from the provided context.

Example

Input

System: Current directory: /home/user
Files in current directory:
- test.txt (empty)
Environment: USER=user, HOME=/home/user
User: ls

Output

test.txt

Intended Use

• Research into learned environment simulation
• Studying command semantics and error modeling
• Dataset generation experiments
• Educational exploration

Limitations

• This model does not execute real commands.
• Even though it's also trained on cases where it's supposed to error out it may still hallucinate incorrect behavior for unseen commands or edge cases.
• File system state must be maintained externally.
• Output accuracy depends heavily on how closely the prompt matches the training format.
• Not suitable for safety-critical or production systems.

Commands It Knows

• pwd
• echo
• cat
• ls
• mkdir
• touch

Any other commands than these might either give wrong results or fail.

Training

• Base model: T5-base
• Fine-tuned on synthetic Linux command datasets
• Mixed-precision training on NVIDIA V100 GPU
• Approximately 80k training samples

Quick Usage

from transformers import T5ForConditionalGeneration, T5Tokenizer
import torch

model_id = "ereniko/LaaLM-v1"

tokenizer = T5Tokenizer.from_pretrained(model_id)
model = T5ForConditionalGeneration.from_pretrained(model_id).cuda()
model.eval()

def run_command(system_state, command):
prompt = f"System: {system_state}\nUser: {command}"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
output = model.generate(
**inputs,
max_new_tokens=128,
do_sample=False
)

return tokenizer.decode(output[0], skip_special_tokens=True)

print(run_command("Current directory: /home/user", "pwd"))

Shoutouts

Very big thanks to @mradermacher for spending time and compute to generate GGUF Quantized versions of LaaLM-v1!

Check out the quantized versions he made here. With these you can run LaaLM-v1 on low tier GPUs or CPUs!