GPT-Flux v1

A domain-specialized Chinese Q&A model for geoscience & flux-tower research

Qwen3-14B · QLoRA fine-tuned on 6,175 QA pairs synthesized from 1,609 peer-reviewed papers

中文简介：通量GPT v1 是基于 Qwen3-14B 的地学/通量站领域问答模型。通过 QLoRA（4-bit）在 1,609 篇地学与涡度相关通量站论文合成的 6,175 条中文问答对上微调 3 epoch 而得，擅长回答碳–水–能量通量、涡度相关方法、能量平衡、蒸散发、遥感尺度转换等专业问题。

Quickstart

from transformers import pipeline

pipe = pipeline("text-generation",
                model="LonghaoWang/flux-gpt-qwen3-14b-merged",
                device_map="auto", torch_dtype="auto")

messages = [{"role": "user", "content": "涡度相关法测量碳通量有哪些主要误差来源？"}]
print(pipe(messages, max_new_tokens=512)[0]["generated_text"][-1]["content"])

Low-VRAM (4-bit) mode — use the LoRA adapter instead (LonghaoWang/flux-gpt-qwen3-14b):

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16)
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B",
    quantization_config=bnb, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "LonghaoWang/flux-gpt-qwen3-14b")

A ready-to-run Colab notebook (free T4 GPU) is in the GitHub repo.

Training

Item	Detail
Base model	`Qwen/Qwen3-14B` (Apache 2.0)
Method	QLoRA — 4-bit NF4, double quant, bf16 compute
LoRA	r=64, α=128, all-linear targets
Data	6,175 Chinese QA pairs (`LonghaoWang/flux-sft`, private)
Epochs / LR	3 / 2e-4 cosine
Seq len	2048 with packing
Batch	2 × 16 grad-accum = 32 effective
Hardware	1× NVIDIA A100 80GB (Hugging Face Jobs)
Train time	~24 min
Final train loss	1.76
GPU cost	~$2.6

Full recipe, scripts and data pipeline: GISWLH/GPT-Flux.

Data

QA pairs were synthesized from the abstract+conclusion chunks of 1,609 peer-reviewed geoscience / eddy-covariance papers using an instruction-tuned LLM, covering flux-tower methods, carbon fluxes (NEE/GPP), energy-balance closure, evapotranspiration, remote sensing & upscaling, and instrumentation QA/QC. See the data card.

Limitations

v1 contains a small number of off-domain papers (e.g. astronomy); a domain filter is planned for v2.
Answers are concise and factual in style; verify critical claims against primary literature.
Inherits the base model's knowledge cutoff and general LLM limitations.

Citation

@software{gptflux2026,
  title  = {GPT-Flux: A Domain-Specialized Chinese Q\&A Model for Geoscience and Flux-Tower Research},
  author = {Wang, Longhao},
  year   = {2026},
  url    = {https://github.com/GISWLH/GPT-Flux}
}

Part of the GPT-Flux collection

🤗 Collection: GPT-Flux
🧩 LoRA adapter: LonghaoWang/flux-gpt-qwen3-14b
📦 Training data: LonghaoWang/flux-sft
💻 Code & docs: GISWLH/GPT-Flux