Gewu-SFT-W8A16

The fine-tuned model of the paper Gewu: Building, Enhancing, and Evaluating a Tool-Augmented 9B Physics Tutoring System (2026) in one file: Qwen/Qwen3.5-9B with the LoRA adapter Alumin-Hydro/Gewu-SFT merged in bfloat16 and then quantised to 8-bit weights with the same recipe as Alumin-Hydro/Gewu-Base-W8A16. Use this file where LoRA hot-mounting is not available; it stands in for the adapter-on model of Gewu's default tier (the two scored level on the hard set, see below); the tier's generic contract and sandbox loop are outside the model (paper Section 4.3). Text-only checkpoint (Qwen3_5ForCausalLM, 32 layers).

Quantisation recipe

llm-compressor GPTQ, scheme W8A16 (int8 weights, 16-bit activations), group size 128, symmetric, static activation order, dampening 0.01, targets every Linear except lm_head, the five projections of each of the 24 linear-attention layers included (248 modules in all); the embeddings, lm_head, the norms and each linear-attention layer's convolution, A_log and dt_bias stay in bfloat16 (the 49 ignore entries in config.json are lm_head and the 24 linear-attention blocks with their norms). Calibration data: 384 problems drawn from GewuBank-20K (seed 0, rendered with the chat template, up to 2,048 tokens), i.e. the model's own working distribution rather than a generic text corpus. Saved in the compressed-tensors pack-quantized format; vLLM 0.27.1 loads it directly. The exact recipe file is included as recipe.yaml.

Measured in the paper

The fine-tuned model in 8 bits passed the paper's pre-registered quantisation gate: +2.5 pp on the hard set against its 16-bit weights, both run on the same RTX 5090 (the 53.6% on the Gewu-SFT card is an A100 run), inside twice the noise floor and not significant (paired p = 0.30), with tool-call rate 59.8% against 58.3% and termination rate 90.4% against 91.2% (paper Section 5.7). The paper's served form is Gewu-Base-W8A16 with the adapter hot-mounted, measured level with this repository's weights on the hard set under the generic contract (−1.4 pp, paired p = 0.60); this repository is the merged alternative.

Serving

vllm serve <local copy of this repo> --served-model-name gewu --dtype bfloat16 --max-model-len 32768 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_xml

License and citation

Part of the adapter's training data (UGPhysics) is CC BY-NC-SA 4.0, so the merged weights are released under CC BY-NC-SA 4.0 (non-commercial, share-alike); the base weights are Apache-2.0. Please 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 seven model files: SHA256SUMS.txt. This card was revised on 2026-10-31; where it differs from the paper, its figures were recomputed from the run logs.