README (test upload)

README.md

@@ -0,0 +1,87 @@
+---
+license: other
+license_name: research-private-mixed
+base_model: autotrust/JEV-9B
+tags:
+  - knowledge-tracing
+  - system-one
+  - lora
+  - private-backup
+---
+
+# jevkt-checkpoints (PRIVATE backup)
+
+Fine-tuned System-One blocks (LoRA adapter + typed decision head + calibration) on top of
+**autotrust/JEV-9B** (Apache-2.0; backbone = Qwen/Qwen3.5-9B), trained in the Jev-KT project
+(repo `codingchild2424/JevKT`, private). This is a backup, not a release. Keep this repo private.
+
+Project documentation: `CLAUDE.md` (session log) and `plan/research_plan.md` in the JevKT repo:
+- E15: plan §9.1 (real KT vs data-free synthetic vs non-KT decision data).
+- E16: plan §9.2 (S1 = new outcome-labelled tasks, S2 = supervised proper scoring vs RL, S3 = RL evidence selection).
+- E17: plan §9.3 (data-free recipes, Y*/O*/D-mix arms).
+- E18 (F2-prior grid D×N×α, SS semi-synthetic), E19 (LLM-student simulator), E20 (MIX arms), LODO A+B+C:
+  CLAUDE.md session log entries 2026-10-12 … 2026-10-19.
+- Backup record: `results/hf_backup.md` in the JevKT repo.
+
+## Layout
+
+```
+e15/<arm>/train_log.jsonl
+e15/<arm>/<ckpt-STEP>/...            # E15 arms (driver jevkt.jev9b_ft)
+e15/F1-cs/<ds>-s<seed>N<N>/ckpt-*/   # E15 F1-cs: one block per cold-start cell
+e16/<arm>/train_log.jsonl
+e16/<arm>/<ckpt-STEP>/...            # trained by the e16 driver (jevkt.e16_rl); also holds E17/E18/E19/E20/LODO arms
+meta_only/pod{A,C}/e1x/<arm>/...     # configs + train logs ONLY (weights still on the stopped pods' volumes)
+```
+
+Each `ckpt-STEP/` contains everything needed to reuse the block:
+- `adapter_vllm/` — vLLM LoRA export (`adapter_config.json`, `adapter_model.safetensors`, `decision_head.json`
+  with head bias / verbalizer ids). Same format as JEV-9B's own `adapter_vllm`.
+- `adapter/` — PEFT LoRA (same weights, PEFT layout).
+- `head.safetensors` — typed decision head; `value_head.safetensors` only for PPO arms (S2-ppo*).
+- `judge_config.json` — slot layout / verbalizers (bare-v1 template); `calibration.json` — per-kind temperature T
+  fitted on the arm's val split.
+No optimizer / resume state was saved by the trainers, so nothing was stripped.
+
+Usage (from the JevKT repo, with JEV-9B served by vLLM):
+`python -m jevkt.jev9b_kt --ds dbe_kt22 --state h25knn --adapter <local ckpt dir> --lora-name <name> --load-lora`.
+`STEP` = number of training rows seen (e.g. ckpt-16000 = 16k-row schedule).
+
+## Arm → training data (paths refer to `LEAP-LAB-KUS/jevkt-data`, i.e. the project's `data/`)
+
+Each `train_log.jsonl` first line records `data`, `data_manifest.train_sha256`, objective, replay file and seed.
+
+| Arm (prefix) | Training data | Notes |
+|---|---|---|
+| e15/F1-LODO-<ds> | data/e15/F1-LODO-<ds> | real KT rows of the other datasets (derived from public KT datasets) |
+| e15/F1-cs/<cell> | data/e15/F1-cs/<cell> | N real learners of the cell (P1 cold-start protocol) |
+| e15/F2-flat | data/e15/F2-flat | data-free F2 prior, flat ranges (synthetic) |
+| e15/F3-gen | data/e15/F3-gen | generic Jev distill corpus (SargeDev/jev-distill-corpus-v3, Apache-2.0); init = fresh |
+| e15/P | data/e15/P | placebo control |
+| e16/S1-ABCLODO16-<ds> | data/e15/F1-pool-jev-LODO-<ds> | A+B+C LODO (real KT of 6 other datasets + Jev teacher labels), 5 % replay |
+| e16/S1-F1pool64-rp | data/e15/F1-pool | A+C (all 7 real KT datasets, in-domain; cold-start caveat in CLAUDE.md 2026-10-12) |
+| e16/S2-<objective> | data/e15/F1-pool | S2: s-brier / grpo / grpo-nostd / rloo / ppo1 / ppo4 on the typed head |
+| e16/S1-G-d<D>-n<N>[-a<α>][-ev100]-syn | data/e18/G-… | E18 data-free F2-prior grid (synthetic) |
+| e16/S1-SS-{hlr,mix}-n16k-syn | data/e18/SS-… | E18 semi-synthetic: real non-KT interaction structure (Duolingo HLR / KuaiRec / Lichess), simulated outcomes |
+| e16/S1-E19-{a1,a2,b}-n<N>[-variant]-syn | data/e19/E19-… | E19 LLM-student simulator (a1 GSM8K+MATH, a2 XES3G5M/KCQRL text, b E17 bank) |
+| e16/S1-MIX-E18E19-n16k-syn, S1-MIX3-n16k-syn | data/e20/MIX… | mixes of E18 / E19 / SS data |
+| Pod A/C arms (meta_only) | data/e15, e16, e17, e18, e19 | see `meta_only/` train logs |
+
+All S1 arms use objective s-log + 5 % replay of `data/e16/replay` (JEV distill corpus subset) unless the train log says otherwise.
+
+## Not in this repo
+
+Arms whose weights are only on the stopped Pod A / Pod C volumes (e.g. E15 F1-pool, F2-syn and training seeds, E17
+Y3x16 / O3x16 / Y*, E18 n4k/n16k grid, A+B+C F1poolJev*, D-seq, T-*) — see `meta_only/` and `results/hf_backup.md`.
+Arms still training at backup time (E19-b-n256k, E19-a2-n16k-a100, S2-grpo-log) are not included.
+
+## Licence caveats
+
+- Base model autotrust/JEV-9B and the replay corpus: Apache-2.0.
+- Many arms were trained on rows **derived from public KT datasets** (ASSISTments 2009/2017, Eedi NIPS34, XES3G5M,
+  DBE-KT22, KDD Cup Algebra 2005, Junyi): E15 F1-*, E16 A+B+C / A+C / LODO / S2. Those datasets carry their own
+  terms (several research-only / non-redistribution). Weights here are research checkpoints; do not release.
+- Some arms contain CC BY-NC data (Duolingo HLR in SS-hlr/SS-mix/T-near/O1-near; ChaosNLI in T-cls/T-mix): any
+  released adapter would inherit NC.
+- Teacher labels come from the commercial TypeSafe Jev API (via OpenRouter); check its terms before any release.
+- Default: **private, internal research use only.**