Gewu-SFT

Gewu-SFT is the fine-tuned model of the paper Gewu: Building, Enhancing, and Evaluating a Tool-Augmented 9B Physics Tutoring System (2026): the base model Qwen/Qwen3.5-9B plus this LoRA adapter (rank 32, alpha 64, twelve target modules including the linear-attention projections; 86,556,672 adapter parameters, 173 MB in bfloat16). In the paper it is the model of the default serving tier; the adapter is attached or detached per request on one weight file.

What it does

Trained on 4,983 of the 5,248 tool-free reasoning traces from a stronger teacher in Alumin-Hydro/gewu-sft-traces (see Training), the adapter shapes behaviour more than it adds physics: the model reasons in Chinese, the students' language, even on most English questions (Chinese on 92.8% of hard-set problems, the base on none), calls a Python sandbox (SymPy, NumPy, SciPy) once the tool is declared at serving time although its training data contains no tool call (on 54.2% of medium-set problems, 137 of 253, with at least one error-free sandbox run on 136 of those 137), is less sensitive to the prompt than the base model, and reaches its answer in roughly a third to a half of the base model's tokens at level or higher accuracy. On the paper's external olympiad set (708 PhysElite items whose key DeepSeek-V4-flash reproduced blind, matched protocol) it scores 62.1% against the base model's 54.5%, and 47.0% against 40.4% on the whole 1,065-item clean pool without that selection; on the paper's medium set it is level with the base (75.5 against 73.1, p = 0.51), as on the paper's random set (77.8 against 78.2, p = 0.92). Details, protocol and noise floors: paper Sections 5.2–5.4.

Training

framework	LLaMA-Factory (0.9.5.dev0); PEFT 0.18.1, Transformers 5.2.0, PyTorch 2.13.0
data	4,983 traces (UGPhysics 2,487 / WebInstruct-verified physics 2,247 / Loong advanced physics 249). Of the 5,248 released, 251 were held out in July for evaluation and 14 were dropped because, with their prompt, they exceeded the 16,384-token cutoff
LoRA	r = 32, alpha = 64, dropout 0.05, targets q/k/v/o, gate/up/down, in_proj_qkv/a/b/z, out_proj
schedule	3 epochs, 936 steps, effective batch 16 (1 x 16 accumulation), peak LR 5e-5, cosine, warmup 0.1
sequence	cutoff 16,384 tokens, thinking kept in the assistant turn, bf16
hardware	one A100 80 GB, 5.9 h

The student was trained, with no system prompt, on the problem plus the 208-character suffix of the paper's generic contract, which asks for a boxed answer and allows \boxed{ILL_POSED} for a defective problem. This contains every instruction the teacher saw. Neither saw the serving system prompt that declares the tool, which is added only at serving time. The training-time chat template is included as chat_template.jinja.

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer
from peft import PeftModel

base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.5-9B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "Alumin-Hydro/Gewu-SFT")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")

The adapter keys follow the layout PEFT produces for the multimodal Qwen3.5 checkpoint (...model.language_model.layers.N...). Stacks that build the text-only module tree (...model.layers.N...), such as AutoModelForCausalLM or vLLM serving Alumin-Hydro/Gewu-Base-W8A16, need the copy in vllm/, which has the language_model. segment removed (PeftModel.from_pretrained(base, "Alumin-Hydro/Gewu-SFT", subfolder="vllm")). vLLM serving the bf16 base builds the multimodal tree and takes the root copy. With the wrong layout PEFT only warns about missing adapter keys, vLLM raises no error, and the model runs as the base. The paper served it merged or hot-mounted under vLLM.

Limitations

A 9B model: on the paper's hard set it answers 53.6% in a single greedy pass, so it is wrong on almost half of those problems; the paper's enhancement tier answers them with eight samples of the base model and a vote, not with this adapter. Under temperature sampling its single samples average 44.2% on the hard set, below its own greedy answer and below the base model's samples (59.5%), so it is a better single answer than a sampler. No 4-bit variant is released: the paper's 4-bit profile failed its pre-registered gate.

License

The base model is Apache-2.0. Part of the training data (UGPhysics) is CC BY-NC-SA 4.0, so this adapter is released under CC BY-NC-SA 4.0 (non-commercial, share-alike). Cite the paper:

Feng, Z. (2026). Gewu: Building, Enhancing, and Evaluating a Tool-Augmented 9B Physics Tutoring System.
Research report, Yau High School Science Award (Computer Science), 2026.

Checksums of the weight, config and template files, in the root and in vllm/, are in SHA256SUMS.txt. This card was revised on 2026-10-31; where it differs from the paper, its figures were recomputed from the run logs and the training files.