README.md

---
library_name: residuals
base_model: google/gemma-3-4b-pt
base_model_relation: adapter
instruct_model: google/gemma-3-4b-it
pipeline_tag: text-generation
tags:
- residuals
- delta
- task-arithmetic
- finetune
---


# Instruction Residuals

This repository contains instruction residuals (delta weights) computed as the parameter-wise difference between `google/gemma-3-4b-it` and `google/gemma-3-4b-pt`.

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


## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from residuals import Residuals

base = AutoModelForCausalLM.from_pretrained("google/gemma-3-4b-pt")
tok = AutoTokenizer.from_pretrained("google/gemma-3-4b-pt")

res = Residuals.from_pretrained("residuals/gemma-3-4b")
res.apply(base, base_tokenizer=tok)
```


## Provenance
- **Created at**: 2025-10-25T18:45:30.975057+00:00
- **DType**: float32
- **Parameters**: 884
- **Shapes hash**: 0aad859058a45de47ddd36c2e67a97e30be0a99b7da51dcfb62e4797e27328d8
- **Names hash**: 22ce2085d7e0c22fd49d378204b2df3f6a4013610f03e7102733ed62b50259c3
- **Base model**: `google/gemma-3-4b-pt`
- **Instruction model**: `google/gemma-3-4b-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