Gemma 4 E4B OpenClaw Agent LoRA
This repository contains a LoRA adapter fine-tuned from unsloth/gemma-4-E4B-it for an OpenClaw-style Python function-calling coding agent.
Training Summary
- Base model:
unsloth/gemma-4-E4B-it
- Dataset:
driaforall/pythonic-function-calling (train)
- Training method: Supervised fine-tuning with Unsloth and TRL
SFTTrainer
- Adapter: LoRA,
r=32, alpha=32, dropout 0
- Tuned modules: language layers, attention modules, and MLP modules; vision layers disabled
- Quantization during training:
load_in_4bit=True
- Max sequence length:
2048
- Batching: per-device batch size
1, gradient accumulation 4
- Learning rate:
0.0002
- Training steps:
1000
- Seed:
3407
Data Formatting
The training data was formatted from driaforall/pythonic-function-calling. Function schemas from dataset system messages were preserved and folded into the first user turn to satisfy Gemma's strict user/model chat alternation. Loss was applied only to model response spans with train_on_responses_only.
Intended Use
This adapter is intended for experiments with Pythonic function-calling and coding-agent behavior. It should be evaluated carefully before use in any production agent loop, especially around tool-call validity and safe code execution.
