GewuBank-20K

GewuBank-20K is a bank of 20,414 physics problems with model-checked keys, arranged on a two-dimensional difficulty grid: a computation type (compute_type; A = analytic, the answer closes in elementary functions; N = numeric, it must be obtained numerically) and a difficulty tier (T1 to T3). It was built for the paper Gewu: Building, Enhancing, and Evaluating a Tool-Augmented 9B Physics Tutoring System (2026), where it is the internal measurement instrument. Every item carries the base model's measurement (Qwen3.5-9B with a Python sandbox offering SymPy, NumPy and SciPy, one greedy pass under the paper's generic contract): whether it reproduced the key, how many tool calls it made, whether it finished, and the sequence length. The rows therefore double as a calibration record, not only as a problem set.

What is in the files

split	rows	what it is
train_pool	19,660	every bank item outside the two frozen evaluation sets
medium_set	253	the paper's medium set (253 items), held out from training by sub-topic; to keep that separation, drop the 1,219 train_pool rows with in_medium_parent = true before training
hard_set	513	the paper's hard set (513 items), held out by item (none of the 513 is in train_pool, but 41 train_pool rows share one of its 60 sub-topics), specified almost entirely at the top tier
contested	419	items whose key the teacher contested (teacher_verdict = wrong); not part of the bank, kept so that the removal is auditable

The bank is the 20,414 rows with in_bank = true (19,660 + 249 + 505). The two evaluation splits are the frozen sets exactly as measured in the paper; the 12 of their items whose key was later contested are flagged in_bank = false rather than removed, so the paper's numbers can be reproduced. Their base_* fields are the calibration measurement (75.1% on the medium set, 54.4% on the hard set), not the reference runs of the paper's Table 4 (73.1 and 57.1). The external evaluation sets of the paper (PhysElite derivatives) are not in this repository; see the paper, Section 3.6.

Columns: uid, problem, gold, unit, answer_type (always NV, one numeric value per key), compute_type (A/N), tier (1 to 3), subject (30 subject codes), subtopic (Chinese sub-topic name, the sub-topic key), subtopic_id (the sub-topic's number within its generation_run and subject, written with an f prefix in the hard set; not a key on its own), generation_run, split, in_bank, in_medium_parent (the 1,495-item medium held-out from which the medium set is drawn), in_random_set (the paper's 500-item random set; all 500 are in train_pool, so drop them before training if you evaluate on it), teacher_verdict (pass / ill_posed / truncated / no_box / wrong / none), teacher_model, verdict_run, and the base-model measurement base_correct (null = no verdict: the numeric checker could not settle the answer and the judge call failed, 58 bank rows), base_tool_calls, base_tool_ok, base_n_tool_error, base_n_turns, base_unfinished, base_budget_exhausted, base_seq_tokens.

How the keys were verified

Items were generated by GPT sol5.6 Ultra from per-sub-topic specifications, with a structural gate at intake (one gradable value per key, tier and computation type declared, deterministic derivation of the key). Every key then received one recorded teacher verdict, from one of three kinds of run (verdict_run): a blind solve by DeepSeek-V4-pro with the sandbox in the distillation runs for the paper's tool-trajectory arms (11,560 items; the Gewu-SFT traces use public problems instead); re-check runs by DeepSeek-V4-flash on 2026-08-28, 09-12 and 09-13 over 8,573 items (A-type items without the tool, N-type items with it), with 246 keys the base model had reproduced as positive controls; and a check of 712 items of the sixth generation batch (v3f) by DeepSeek-V4-pro and V4-flash, with the tool on every item. Verdicts are recorded per item and no key was altered. Only the 431 items the teacher marked wrong were removed from the bank (the 419 rows of the contested split and the 12 flagged rows of the evaluation splits); items with any other verdict stay. Among the bank's rows the verdicts are: pass 20,072, ill_posed 153, truncated 110, no_box 79, none 0. Full protocol: paper Sections 3.2 to 3.4 and Appendix D.

Base-model measurement

Qwen3.5-9B with the Python sandbox (SymPy, NumPy, SciPy), generic contract, greedy, 16,384 tokens per turn, 32,768 tokens for the whole sequence, at most twelve tool turns. The deterministic numeric-equivalence checker credits the answers it finds equivalent, and every other answer goes to a model judge, which settled 5,186 of the 20,414 bank items (256 of them credited). In two calibration slices (985 bank items) every judge call failed, so there 248 answers stay marked wrong by the checker alone and 58 have no verdict. On the graded bank items the base model reproduces the key on 74.56% (15,178 of 20,356); the paper's Figure 2 gives the grid (its six cells count the 58 ungraded items as wrong).

Intended use and caveats

Training and evaluating physics solvers; studying tool use on numeric problems. The keys were verified by models, not by a human panel; the paper reports the audit rates. The problems are in Chinese with LaTeX. Noise floors: an accuracy on the hard set carries a 1-sigma binomial floor of about 2.2 pp, on the medium set 2.7 pp.

License and citation

Released under CC BY 4.0. 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.

Companion repositories: Alumin-Hydro/gewu-sft-traces (the distillation traces) and Alumin-Hydro/Gewu-SFT (the fine-tuned adapter). Checksums of the data files are in SHA256SUMS.txt. This card was revised on 2026-10-31; where it differs from the paper, its figures were recomputed from the data files and the run logs.