bashgemma-270m

BashGemma 270M

Fine-tuned FunctionGemma 270M for natural language to bash command translation.

Paper: Zenodo DOI: 10.5281/zenodo.18058613

Model Description

BashGemma translates natural language queries into structured JSON tool calls representing bash commands. It achieves 57.4% NLC2CMD accuracy on NL2Bash test data—a 52.9 percentage point improvement over the base FunctionGemma model.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load model
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/functiongemma-270m-it",
    torch_dtype=torch.float32,
)
model = PeftModel.from_pretrained(base_model, "thinkthink-dev/bashgemma-270m")
tokenizer = AutoTokenizer.from_pretrained("thinkthink-dev/bashgemma-270m")

# Generate
prompt = "<start_of_turn>user\nFind all Python files in the current directory<end_of_turn>\n<start_of_turn>model\n"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=150, do_sample=False)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))
# {"name": "find", "arguments": {"name": "'*.py'", "path": "."}}

Capabilities

Strong performance on:

• find with name, type, size, mtime filters
• Simple pipelines (2-3 commands)
• Common utilities: grep, ls, cat, wc, sort

Limitations:

• Complex multi-step operations
• Utilities not in training data (rsync, tar, etc.)
• Complex sed/awk patterns

Training

• Base Model: unsloth/functiongemma-270m-it
• Method: LoRA (rank=64, alpha=128)
• Training Data: 9,153 NL2Bash examples
• Training Time: 36 minutes on Apple M4 Max
• Approach: Response-only training (mask instruction tokens)

Citation

@software{large2024bashgemma,
  author = {Large, Jack},
  title = {BashGemma: Fine-tuning a 270M Parameter Model for Natural Language to Bash Translation},
  year = {2024},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18058613},
  url = {https://zenodo.org/records/18058613}
}

License

Apache 2.0 (same as base Gemma model)