ensembleai/Qwen3-1.7B-CSQA-NdLinearLoRA

NdLinear-LoRA Fine-Tuned Models

This repository contains a collection of language models fine-tuned using a custom NdLinear-LoRA architecture. NdLinear-LoRA is a variant of Low-Rank Adaptation (LoRA) that reshapes weight matrices into N-dimensional tensors and applies a factorized linear transformation for parameter-efficient fine-tuning.

Available Models

Below is a list of the fine-tuned models. For best results, it's recommended to host each model in its own repository on the Hugging Face Hub.

Fine-Tuned Model Name	Base Model	Fine-Tuning Dataset
Meta-Llama-3-8B-CSQA-NdLinearLoRA	meta-llama/Llama-3-8B	commonsense_qa
Meta-Llama-3-8B-Math10K-NdLinearLoRA	meta-llama/Llama-3-8B	lmms-lab/Math10K
Qwen3-1.7B-CSQA-NdLinearLoRA	Qwen/Qwen3-1.7B-Base	commonsense_qa
Qwen3-1.7B-Math10K-NdLinearLoRA	Qwen/Qwen3-1.7B-Base	lmms-lab/Math10K

How to Use

Because these models use a custom architecture, you must pass trust_remote_code=True when loading them. This allows the transformers library to download and use the modeling_ndlinear.py file that should be included in each model's repository.

Dependencies: Before you start, make sure you have the necessary libraries installed:

pip install torch transformers safetensors huggingface_hub accelerate
pip install ndlinear

Example Loading Script

This script will work for any of the models listed above. Just change the REPO_ID.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# --- Example Usage ---

# 1. Choose the model you want to use from the table above
#    Replace "YourUsername" with your Hugging Face username or organization.
REPO_ID = "YourUsername/Qwen3-1.7B-Math10K-NdLinearLoRA" 

# 2. Load the model and tokenizer
# `trust_remote_code=True` is required to load the custom architecture.
print(f"Loading model: {REPO_ID}")
model = AutoModelForCausalLM.from_pretrained(
    REPO_ID,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)
print("Model and tokenizer loaded successfully.")


# 3. Generate text
# This prompt is geared for a math model. Adjust it for a QA model if needed.
prompt = "### Instruction:\\nSolve the following math problem: If a train travels at 60 miles per hour, how long does it take to travel 180 miles?\\n\\n### Solution:\\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=150, eos_token_id=tokenizer.eos_token_id)

print("\\n--- Generated Output ---")
print(tokenizer.decode(outputs[0], skip_special_tokens=True))