Hugging Face
Search models, datasets, users...
Ctrl K
Models
Datasets
Spaces
Buckets
Docs
Enterprise
Pricing
Website
Tasks
HuggingChat
Collections
Languages
Organizations
Community
Blog
Posts
Daily Papers
Hardware
Learn
Discord
Forum
GitHub
Solutions
Team & Enterprise
Hugging Face PRO
Enterprise Support
Inference Providers
Inference Endpoints
Storage Buckets
Log In
Sign Up
Spaces:
cazyundee
/
training
Like
0
Running
on
Zero
App
Files
Files
Community
Fetching metadata from the HF Docker repository...
main
training
/
tinychess
243 kB
Ctrl+K
Ctrl+K
2 contributors
History:
28 commits
cazyundee
maxk config (NEW MANDATE: 100K budget): baseline arch at d72/dff160/d_attn48/d_move36 = 99,374 params - the 2000-Elo chassis absorbing all 18 proven feature channels + jitter; forward-verified
84b2a92
1 day ago
__pycache__
fastwide config: 200,206-param recurrent control for the transformer comparison (param-matched within 0.4%); ignore __pycache__
3 days ago
__init__.py
Safe
0 Bytes
tinychess: self-play research substrate (phase 1-4)
2 months ago
arena.py
Safe
6.86 kB
tinychess: self-play research substrate (phase 1-4)
2 months ago
batching.py
Safe
3.79 kB
fix: repetition feature lost in FEN round-trip (train/inference mismatch); greedy-health exp
about 2 months ago
checkpoint.py
Safe
3.49 kB
load_checkpoint zero-pad shim (the resume path lacked it - C3 FATAL'd loading the 12ch c1 ckpt under the 18ch encoding; append-only feature compatibility now holds for BOTH load paths; verified: c1@1200 resumes at step 1200)
1 day ago
config.py
Safe
4.25 kB
maxk config (NEW MANDATE: 100K budget): baseline arch at d72/dff160/d_attn48/d_move36 = 99,374 params - the 2000-Elo chassis absorbing all 18 proven feature channels + jitter; forward-verified
1 day ago
encoding.py
Safe
13 kB
Option K: king-safety input channels (6 rules-derived SQ_EXTRA: own/opp king rings + shields + enemy-attacks-in-ring danger map + mate-threat band) targeting the SF-loss diagnosis (0/70 castled, mated at parity); +576 params (201,550); spot-verified (rings exact, shields correct, Qh4->f2 danger lights); c1@1200 loads through 18ch encode via shim; encode 0.74ms/pos
1 day ago
model.py
Safe
26.4 kB
step_emb_jitter: train-time random step-row sampling so untrained depth rows get gradient
3 days ago
parallel.py
Safe
6.93 kB
spawn-based persistent self-play pool (fork/libgomp deadlock fix)
about 2 months ago
plasticity.py
Safe
9.57 kB
tinychess: full research artifact - self-play RL, ablations, compute scaling, faithfulness, precision, plasticity
2 months ago
quant.py
Safe
2.78 kB
tinychess: self-play research substrate (phase 1-4)
2 months ago
replay.py
Safe
11.1 kB
s7 halt-alignment: --halt-align aux BCE(p_halt, 1[conv<median]) with play-time conv |V_t+1-V_t| stored per-sample; legacy paths untouched; smokes green
2 days ago
selfplay.py
Safe
22.5 kB
K3 curriculum: castle_danger generator - near-start positions with FULL castle rights + attacker array aimed at the king (fixes K2's two documented deficits: rights and punishment co-present); 60/60 valid with rights+danger, castling legal in 18% at ply 0 (rights persist); mixed_castle 50/50 mode
1 day ago
thought.py
Safe
8.98 kB
tinychess: full research artifact - self-play RL, ablations, compute scaling, faithfulness, precision, plasticity
2 months ago