residuals/gemma-3-12b

Instruction Residuals

This repository contains instruction residuals (delta weights) computed as the parameter-wise difference between google/gemma-3-12b-it and google/gemma-3-12b-pt.

Apply these residuals to the base model to reconstruct the instruction-tuned weights without retraining.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from residuals import Residuals

base = AutoModelForCausalLM.from_pretrained("google/gemma-3-12b-pt")
tok = AutoTokenizer.from_pretrained("google/gemma-3-12b-pt")

res = Residuals.from_pretrained("residuals/gemma-3-12b")
res.apply(base, base_tokenizer=tok)

Provenance

• Created at: 2025-10-25T18:51:41.834187+00:00
• DType: float32
• Parameters: 1066
• Shapes hash: 9c53e73a967d7ab81405b64f3a4d8f48f2a09ebac740aba010a71abec2548555
• Names hash: cc9f94a44626d0ab23b382b46a5f56203697a18e8f408301387e56c8a81aea51
• Base model: google/gemma-3-12b-pt
• Instruction model: google/gemma-3-12b-it

Files

• model.safetensors: Serialized residual tensors (safetensors format).
• (optional) model.safetensors.index.json + shard files model-00001-of-000N.safetensors, ... for multi-part weights.
• config.json: Residuals metadata and provenance.
• tokenizer files: Saved tokenizer for compatibility.

About this format

These are additive residuals (task vectors). Applying them to the base model's parameters reconstructs the instruction-tuned model.

Tools

Generated with the residuals Python package. Install via: pip install residuals.

• PyPI: https://pypi.org/project/residuals/
• Source: https://github.com/omarish/residuals