Unggi commited on
Commit
3c9c77e
·
verified ·
1 Parent(s): dcb97b4

README: add pykt/classical, final layout

Browse files
Files changed (1) hide show
  1. README.md +8 -0
README.md CHANGED
@@ -32,8 +32,15 @@ e15/F1-cs/<ds>-s<seed>N<N>/ckpt-*/ # E15 F1-cs: one block per cold-start cell
32
  e16/<arm>/train_log.jsonl
33
  e16/<arm>/<ckpt-STEP>/... # trained by the e16 driver (jevkt.e16_rl); also holds E17/E18/E19/E20/LODO arms
34
  meta_only/pod{A,C}/e1x/<arm>/... # configs + train logs ONLY (weights still on the stopped pods' volumes)
 
 
 
35
  ```
36
 
 
 
 
 
37
  Each `ckpt-STEP/` contains everything needed to reuse the block:
38
  - `adapter_vllm/` — vLLM LoRA export (`adapter_config.json`, `adapter_model.safetensors`, `decision_head.json`
39
  with head bias / verbalizer ids). Same format as JEV-9B's own `adapter_vllm`.
@@ -65,6 +72,7 @@ Each `train_log.jsonl` first line records `data`, `data_manifest.train_sha256`,
65
  | e16/S1-SS-{hlr,mix}-n16k-syn | data/e18/SS-… | E18 semi-synthetic: real non-KT interaction structure (Duolingo HLR / KuaiRec / Lichess), simulated outcomes |
66
  | 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) |
67
  | e16/S1-MIX-E18E19-n16k-syn, S1-MIX3-n16k-syn | data/e20/MIX… | mixes of E18 / E19 / SS data |
 
68
  | Pod A/C arms (meta_only) | data/e15, e16, e17, e18, e19 | see `meta_only/` train logs |
69
 
70
  All S1 arms use objective s-log + 5 % replay of `data/e16/replay` (JEV distill corpus subset) unless the train log says otherwise.
 
32
  e16/<arm>/train_log.jsonl
33
  e16/<arm>/<ckpt-STEP>/... # trained by the e16 driver (jevkt.e16_rl); also holds E17/E18/E19/E20/LODO arms
34
  meta_only/pod{A,C}/e1x/<arm>/... # configs + train logs ONLY (weights still on the stopped pods' volumes)
35
+ pykt/<ds>/pykt-<model>-s<seed>/ # pyKT DLKT baselines (local machine copy): qid_model.ckpt + train_info.json;
36
+ pykt/<ds>/coldstart/... # full-data seeds + cold-start cells (jevkt.pykt_runner; CLAUDE.md 2026-10-03 … 10-07)
37
+ classical/<model>__<ds>__s<seed>.pt # leap-kt-toolkit baselines (src/jevkt/classical)
38
  ```
39
 
40
+ pyKT note: this is the LOCAL machine's `models/pykt` (19.8 GB). Pod B's `models/pykt` (8.5 GB: Pod-B-trained Stage-3
41
+ seeds, cold-start cells, tuned DLKT) and Pod A's pyKT checkpoints are not included; predictions for all of them
42
+ are in the JevKT repo `results/preds/`.
43
+
44
  Each `ckpt-STEP/` contains everything needed to reuse the block:
45
  - `adapter_vllm/` — vLLM LoRA export (`adapter_config.json`, `adapter_model.safetensors`, `decision_head.json`
46
  with head bias / verbalizer ids). Same format as JEV-9B's own `adapter_vllm`.
 
72
  | e16/S1-SS-{hlr,mix}-n16k-syn | data/e18/SS-… | E18 semi-synthetic: real non-KT interaction structure (Duolingo HLR / KuaiRec / Lichess), simulated outcomes |
73
  | 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) |
74
  | e16/S1-MIX-E18E19-n16k-syn, S1-MIX3-n16k-syn | data/e20/MIX… | mixes of E18 / E19 / SS data |
75
+ | pykt/, classical/ | public KT datasets (data/<ds>, not in jevkt-data) | DLKT baselines |
76
  | Pod A/C arms (meta_only) | data/e15, e16, e17, e18, e19 | see `meta_only/` train logs |
77
 
78
  All S1 arms use objective s-log + 5 % replay of `data/e16/replay` (JEV distill corpus subset) unless the train log says otherwise.