import os
import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer, BitsAndBytesConfig

token = os.environ["HF_TOKEN"]
base_id = "google/gemma-4-E2B-it"
adapter_id = "nabin2004/AOS-gemma4-manim-sft"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

base = AutoModelForImageTextToText.from_pretrained(
    base_id,
    quantization_config=bnb,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    token=token,
)
model = PeftModel.from_pretrained(base, adapter_id, token=token)
tokenizer = AutoTokenizer.from_pretrained(adapter_id, token=token)
model.eval()

Inference with AOS tool loop

The adapter is trained on multi-turn Code Agent tool calls, not single-turn prose. Use the AOS inference script:

bash

cd apps/sft
uv run python infer.py \
  --adapter-dir nabin2004/AOS-gemma4-manim-sft \
  --prompt "Create a short Manim scene explaining eigenvectors in 2D."

Colab:

bash

uv run --package sft python apps/sft/infer.py \
  --adapter-dir nabin2004/AOS-gemma4-manim-sft \
  --colab \
  --prompt "Create a short Manim scene explaining eigenvectors in 2D."

Files

File	Description
`adapter_config.json`	PEFT LoRA configuration
`adapter_model.safetensors`	LoRA weights
`tokenizer_config.json`	Training chat template (includes `{% generation %}` markers)

Training

bash

cd apps/sft
uv run python run.py --colab --epochs 1 --report-to none --push-to-hub

Or upload an existing adapter:

bash

export HF_TOKEN=hf_...
uv run python upload_adapter.py --adapter-dir ./gemma4-manim-ft

AOS-Trajectories dataset
ManiBench — GRPO benchmark (Phase 2)
AOS repository

nabin2004/AOS-gemma4-manim-sft

AOS Gemma 4 Manim SFT (LoRA)

Base model

Training data

Usage

Load with PEFT (Python)

Inference with AOS tool loop

Files

Training

Related

nabin2004/
AOS-gemma4-manim-sft