jevkt-synthetic

Synthetic training corpora of the Jev-KT project (knowledge tracing with System-One typed-decision models; LoRA checkpoints in LEAP-LAB-KUS/jevkt-checkpoints). This repo holds only data with no real KT-benchmark learner sequences and no non-commercial-licence source content. Real-dataset-derived rows, CC BY-NC sources and unclear-licence material are kept in a separate private repo.

Every arm directory has manifest.json (builder sha256, train/val sha256, counts, sources and licences) plus train.jsonl / val.jsonl in the trainer's format (text = serialised [kind] [state] [question] [options] request, target = probability vector, y = outcome where defined). Each top-level directory has its own README naming the generator and the label source.

Path	Content	Licence
e15/	F2-syn (wide prior), F2-flat, F3-gen, F4-mix	Apache-2.0 (F3/F4 rows from SargeDev/jev-distill-corpus-v3, Apache-2.0)
e16/	T-seq, P-seq, D-seq, T-fc (Lichess, KuaiRec, football results, ForecastBench), replay	CC BY-SA 4.0 (replay: Apache-2.0)
e17/	data-free recipes Y1/Y1b/Y2/Y3/Y3G/Y4-t*, generated general tasks G/_g, Jev teacher files	Apache-2.0 + provider terms
e18/	F2-prior grid G-d-n[-a<α>][-ev100], task store, Jev teacher files	Apache-2.0
e19/	LLM-student simulator, pools a1 (GSM8K + MATH, MIT) and b (E17 bank)	Apache-2.0 + MIT item text + student-model terms
e20/	MIX-E18E19, TA-G512, TA-E19a1	Apache-2.0
synth_bank/, synth_e17/, synth_s/, synth_l/	generated item banks, statistical / LLM-agent synthetic learners	Apache-2.0 + provider terms
synth_prior/	F2-prior task corpora f2_wide_s0, f2_flat_s0, e18_s0	Apache-2.0
xover_*/	six crossover synthetic datasets (BKT + IRT simulator)	Apache-2.0
backups/	paid-output archives for E19 pool b and soft labels	as e19

Teacher labels (important)

Many targets mix a simulator / outcome label with a commercial Jev label: TypeSafe Jev called through OpenRouter's System-One endpoint. Versions: e16 D-seq, e17, e18 = typesafe/jev-1.13-20260917; e19 = typesafe/jev-1.14-20261014 (the E19 manifests' "jev-1.13" string is a known mislabel). The response files keep the model id per call. Jev is closed and not deterministic (P varies about ±0.03 across identical calls).

Third-party terms

• LLM-generated items and simulated student answers come from openai/gpt-4o-mini, meta-llama/llama-3.3-70b, openai/gpt-oss-120b (verifiers), and the E19 students (Llama 3.1 / 3.2, Gemma 3, Qwen 2.5). Their providers' terms (OpenAI usage policies, Llama Community Licence, Gemma Terms of Use) may restrict how models trained on these outputs can be named or released.
• ShareAlike: derivatives of e16/ must stay CC BY-SA 4.0.

Not included (private)

Rows built from real KT benchmarks (ASSIST09/17, NIPS34, XES3G5M, DBE-KT22, Algebra 2005, Junyi), CC BY-NC sources (Duolingo HLR, ChaosNLI) and the E19 a2 pool (XES3G5M question text via KCQRL, licence unclear).