boomerang-qwen3-2.3B

Model Description

Boomerang distillation is a phenomenon in LLMs that allows us to distill a teacher model into a student and reincorporate teacher layers to create intermediate-sized models with no additional training. This is the student model distilled from Qwen3-4B-Base from our paper.

Training Procedure

This model was initialized from Qwen3-4B-Base by copying every other layer and the last 2 layers. It was distilled on 2.1B tokens of The Pile deduplicated with cross entropy, KL, and cosine loss to match the activations of Qwen3-4B-Base. We used the following hyperparameters:

• Learning rate: 3e-4
• Learning rate scheduler: cosine
• Warmup ratio: 0.01
• Optimizer: AdamW
• Adam betas: (0.9, 0.95)
• Adam epsilon: 1e-8
• Weight decay: 0.1
• Max. gradient norm: 1.0
• Number of training steps: 500
• Max. sequence length: 2048
• Effective batch size: 2048
• Mixed precision: bf16
• KLDiv weight: 0.1
• Cosine distance weight per layer: 0.1

Use

To interpolate between this model and Qwen3-4B-Base, please use the build_intermediate_model function from our github repository:

import torch
from patching.patch import build_intermediate_model

intermediate_model = build_intermediate_model(
  teacher_name_or_path = "Qwen/Qwen3-4B-Base",
  student_name_or_path = "Harvard-DCML/boomerang-qwen3-2.3B",
  num_layers_to_patch = 2,
  patch_first_k_layers = False,
  dtype = torch.bfloat16,
)

Notes:

1. Changing num_layers_to_patch changes the size of the intermediate model by patching different numbers of student layers.
2. patch_first_k_layers should be set to False for this model for optimal interpolation performance.

Citation

@article{kangaslahti2025boomerang,
  title={Boomerang Distillation Enables Zero-Shot Model Size Interpolation},
  author={Kangaslahti, Sara and Nayak, Nihal V and Geuter, Jonathan and Fumero, Marco and Locatello, Francesco and Alvarez-Melis, David},
  journal={arXiv preprint arXiv:2510.05064},
  year={2025},
  url={https://arxiv.org/abs/2510.05064}
}