homeGPT
GPT-2 Shared Layer Reconstructor — a lightweight trained module that reconstructs shared transformer layers on top of a frozen GPT-2 base model (the frozen GPT-2 weights are downloaded automatically at load time).
- Base model: openai-community/gpt2
- Tensor type: F32
- Task: Feature Extraction
Usage
High-level pipeline
from transformers import pipeline
pipe = pipeline(
"feature-extraction",
model="AIhomeJP/homeGPT",
trust_remote_code=True,
device_map="auto",
)
Load directly
from transformers import AutoModel
model = AutoModel.from_pretrained(
"AIhomeJP/homeGPT",
trust_remote_code=True,
dtype="auto",
device_map="auto",
)
Load the reconstructor block only
from transformers import AutoModel
reconstructor = AutoModel.from_pretrained(
"AIhomeJP/homeGPT",
trust_remote_code=True,
)
Assemble the runtime student
Use build_student_from_pretrained() from the training script to attach the frozen GPT-2 token/position embeddings, final LayerNorm, and LM head to this reconstructor.
Model Details
This repository contains only the trained shared-layer reconstructor weights. At inference time, the frozen GPT-2 base components (token/position embeddings, final LayerNorm, LM head) are fetched from openai-community/gpt2 and combined with the reconstructor to form the full runtime model.
License
MIT