avewright commited on
Commit
768cbc8
·
verified ·
1 Parent(s): a294c7c

Upload export_summary_2026-03-31.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. export_summary_2026-03-31.md +69 -0
export_summary_2026-03-31.md ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Export Summary 2026-03-31
2
+
3
+ ## Selected Model
4
+
5
+ - Hugging Face model repo: `avewright/chess-transformer-200m-latest`
6
+ - Selected remote revision for republishing as `latest_model.pt`: `c01d9434fbcd1a5e501f2f82ff88e89134c30f29`
7
+ - Selected checkpoint status:
8
+ - `epoch=7`
9
+ - `train_steps=2742`
10
+ - `dataset_records=107851`
11
+ - `last_eval.loss=1.3680388554930687`
12
+ - `last_eval.acc=0.40478515625`
13
+ - `last_eval.top3=0.69970703125`
14
+ - Reason selected:
15
+ - quick Elo test `outputs/elo_eval_paused_exp084_current.json` estimated `1646`
16
+ - later quick Elo test `outputs/elo_eval_paused_exp084_current2.json` estimated `1600`
17
+ - this earlier checkpoint was the stronger tested model
18
+
19
+ ## Frozen Dataset
20
+
21
+ - Hugging Face dataset repo target: `avewright/exp085-parallel-multipv-harvest`
22
+ - Local dataset source: `outputs/exp085_parallel_multipv_harvest`
23
+ - Frozen after stopping live jobs and checkpointing SQLite WAL
24
+ - Canonical frozen counts from JSONL line counts:
25
+ - `dataset_files=44`
26
+ - `dataset_records=224191`
27
+ - `dataset_bytes=551330268`
28
+ - `last_shard=positions_000044.jsonl`
29
+ - `last_shard_records=2724`
30
+ - Final status snapshot:
31
+ - `written_positions=217718`
32
+ - `exact_positions_seen=224181`
33
+ - `records_per_min=553.31`
34
+ - Note:
35
+ - JSONL line counts are treated as canonical for export because they represent the durable training corpus on disk
36
+
37
+ ## Runtime State At Freeze
38
+
39
+ - Training process: stopped
40
+ - Harvester process: stopped
41
+ - `tmux` sessions: stopped
42
+ - Local resumable trainer checkpoint retained at `outputs/exp084_old_model_on_exp083/checkpoints/latest.pt`
43
+
44
+ ## Artifacts Published
45
+
46
+ - Model repo:
47
+ - `latest_model.pt`
48
+ - `status.json`
49
+ - `selected_elo_model_step2742.pt`
50
+ - `selected_elo_model_step2742_status.json`
51
+ - `selected_elo_model_step2742_summary.json`
52
+ - Elo logs/json for both quick tests
53
+ - Dataset repo:
54
+ - `README.md`
55
+ - `manifest.json`
56
+ - `status.json`
57
+ - `exp085.log`
58
+ - `stdout.log`
59
+ - `seen_positions.sqlite`
60
+ - `dataset/positions_*.jsonl`
61
+
62
+ ## GitHub Push
63
+
64
+ - Git remote: `origin https://github.com/avewright/transform.git`
65
+ - Branch: `main`
66
+ - Repository push includes:
67
+ - code changes
68
+ - architecture doc updates
69
+ - export summary and dataset card