Harvard-DCML

ADAPT-Olmo3-4.3B-Instruct

README

License: apache-2.0

Model Description

ADAPT is a technique that allows for size interpolation across different post-trained variants of the same base model. This is the student model distilled from Olmo-3-7B-Instruct from our paper.

Training Procedure

This model was initialized from Olmo-3-7B-Instruct by copying every other layer and the last 2 layers. It was distilled on 1B tokens of The deduplicated Pile and 1B of the math split from Llama Nemotron Post Training Dataset with cross entropy, KL, and cosine loss to match the activations of Olmo-3-7B-Instruct. 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 (Pre-training phase): 480
Effective batch size (Pre-training phase): 2048
Number of training steps (SFT phase): 585
Effective batch size (SFT phase): 4096
Max. sequence length: 1024
Mixed precision: bf16
KLDiv weight: 0.1
Cosine distance weight per layer: 10.0

Use

To interpolate between this model and Olmo-3-7B-Instruct, please use the build_intermediate_model function from our github repository:

python3
import torch
from patching.patch import build_intermediate_model

intermediate_model = build_intermediate_model(
  teacher_name_or_path = "allenai/Olmo-3-7B-Instruct",
  student_name_or_path = "Harvard-DCML/ADAPT-Olmo3-4.3B-Instruct",
  num_layers_to_patch = 2,
  patch_first_k_layers = True,
  dtype = torch.bfloat16,
)

Notes:

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

Citation

markdown
@misc{zhou2026thinkingrightsizeamortized,
      title={Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs}, 
      author={Yan Zhou and Sara Kangaslahti and Jonathan Geuter and Nihal V. Nayak and Marco Fumero and Francesco Locatello and David Alvarez-Melis},
      year={2026},
      eprint={2608.22854},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.22854}, 
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Harvard-DCML

Model Tree

Base

allenai/Olmo-3-7B-Instruct

Fine-tuned

this model

Input Modalities

Text

Output Modalities