🧠 Neur-0.0-Full

Fine-tuned variant of openai/gpt-oss-20b using supervised fine-tuning (SFT) with QLoRA adapters, later merged into a full standalone model.

---

📋 Overview

Neur n0.0 is a 20B-parameter transformer language model derived from GPT-OSS-20B, fine-tuned to improve reasoning, coding, and multi-step tool-use behaviors.

n0.0 is intended to be used in conjunction with Agentic tools for the purpose of helping the user complete everyday tasks on the computer.

It can also directly write or edit code on any IDE or platform at the user's request.

This release merges LoRA adapters into the base model so it can be loaded directly with from_pretrained()—no external adapter weights required.

---

🔧 Model Details

Attribute	Value
Base Model	openai/gpt-oss-20b
Architecture	Decoder-only Transformer (GPT-style)
Parameters	~20.9B
Precision	bfloat16 (BF16)
Fine-tuning Method	QLoRA (4-bit quantization, r=16, α=32)
Dataset	CodeAlpaca-20k + curated reasoning data
Training Steps	1500
Optimizer	AdamW with cosine LR schedule
Output Directory	out-sft-qlora/merged-standalone
Frameworks	🤗 Transformers, PEFT, BitsAndBytes, Accelerate

---

🚀 Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "xenon111/neur-0.0-full"

tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype=torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported()
           else (torch.float16 if torch.cuda.is_available() else torch.float32),
    trust_remote_code=True,
)

prompt = "Write a Python function that sorts a list of numbers using merge sort."
inputs = tok(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=200, temperature=0.7)
print(tok.decode(output[0], skip_special_tokens=True))