TCGBaselines/deep_encoders

Current deep specialist encoder leaderboard

This repository contains the complete, checksum-bound JAX checkpoint set referenced by the canonical configs/train_leaderboard.json snapshot.

• Source code Git commit: 2d993bc0c770396da69e4227c67696751beba8ff
• Packaged worktree board Git blob: f2e5cacc9d3fe203b6f7470b435c1b5b89cb4326
• Board blob at source commit: 5cf2495637ba8bcbdd192e85d0ca61aba2c62ff3
• Packaged board matches source commit: false
• Source board SHA-256: 3b94e5d10247513d3f4d74ace1cac39776def63b56721fc8f5451fbefbcba364
• Board members / unique deck hashes: 20 / 20
• Unique checkpoint directories: 19
• Direct deck-trained members: 17

Every payload is a complete entity_interaction_deep, stateless_mlp, specialist_rule_state_v1, public-event-window-0 checkpoint. The archive retains every file in each referenced checkpoint directory and deduplicates directories reused by more than one board member.

Download, verify, and extract

hf download TCGBaselines/deep_encoders \
  deep_encoders_train_leaderboard_current.tar.gz deep_encoders_train_leaderboard_current.tar.gz.sha256 --local-dir .
sha256sum -c deep_encoders_train_leaderboard_current.tar.gz.sha256
tar -xzf deep_encoders_train_leaderboard_current.tar.gz
sha256sum -c deep_encoders_bundle_checksums.sha256

Extract at the ladderboard_elo repository root. This creates the exact checkpoints/<id>/ paths plus configs/train_leaderboard.json, MODELS.json, DEEP_ENCODERS_BUNDLE_README.md, and the internal checksum manifest. Existing files at those paths are replaced by normal tar extraction, so use a clean checkout or back up local changes first.

Cross-deck actor reuse

board_deck_hash and checkpoint_trained_deck_hash are separate fields in MODELS.json. The following entries intentionally run an actor on a target deck other than the deck recorded by its immutable checkpoint manifest:

Member	Board deck	Checkpoint-trained deck	Relation
deep_blaziken_dragapult_phase2_u000656_aa65cb	362eca0b8a31393e	9ed6ec2254b54b9b	deck_revision_actor_retarget
raging_bolt_source_actor_baseline_u0	ed36260904773f9d	9ed6ec2254b54b9b	source_actor_baseline_via_deck_revision_retarget
okidogi_source_actor_baseline_u0	b15fcfd19acc20ed	8e33dea1649f0c39	source_actor_baseline

These are valid frozen actor opponents, but their stored critics are not declared compatible with the retargeted deck. Use actor-only initialization and a fresh/warm-started target-deck critic unless newer deck-specific evidence explicitly says otherwise. The immutable checkpoint manifests are preserved rather than rewritten to conceal training provenance.

Current board members

Rank	Elo	Member	Board deck	Checkpoint relation
1	1049.5	festival_lead_plain_ab_best_u020	07d4c9cb1496190b	direct
2	1030.4	alakazam_dudunsparce_phase1_u000160_2f6942	01dc0c8ee2afb65c	direct
3	1012.7	lopbunny_dudunsparce_phase1_u000240_71df17	04689d80f09215b6	direct
4	989.8	cynthia_garchomp_phase1_u000160_64d81c	0f40efbf19fcd898	direct
5	959.1	mega_lucario_phase1_u000160_00c891	2796018749350c2b	direct
6	958.1	prague_dragapult_phase1_u000160_0f81ba	9d406aa686b18487	direct
7	947.9	archaludon_cinderace_phase1_u000160_e77531	82ce67869190aefd	direct
8	946.2	deep_mega_starmie_cinderace_phase2_u000196_db88e4	00bbd953f688a18b	direct
9	941.3	deep_blaziken_dragapult_phase2_u000656_aa65cb	362eca0b8a31393e	deck_revision_actor_retarget
10	941.3	raging_bolt_source_actor_baseline_u0	ed36260904773f9d	source_actor_baseline_via_deck_revision_retarget
11	940.3	hop_trevenant_dudunsparce_deep_vrpo_final_u000740	850f423d1fe52b95	direct
12	938.9	tr_honchkrow_porygon2_deep_vrpo_phase2_u000603_e854ed	d71c667d3ed25c89	direct
13	927.3	starmie_froslass_phase2_u000925_efff34	0e6ead751f2c6e11	direct
14	927.0	tr_mewtwo_spidops_deep_vrpo_u000480_620d59	ec38a9f1af1ad0f3	direct
15	919.9	dragapult_dusknoir_phase1_u000220_8e22f9	f9f5dc7482675db1	direct
16	916.8	lucario_dunsparce_deep_vrpo_continuation_u000760_1c1db6	8e33dea1649f0c39	direct
17	913.2	crustle_kangaskhan_deep_ppo_phase2_u000121_b43489	2959276a9ee66220	direct
18	904.4	okidogi_source_actor_baseline_u0	b15fcfd19acc20ed	source_actor_baseline
19	877.1	grimmsnarl_froslass_deep_vrpo_phase2_u000278_0fd730	2a9b67c9561fbdd4	direct
20	816.2	mega_abomasnow_kyogre_deep_vrpo_u000400_774a80	b5a0ec95d155c444	direct

These are game-playing specialist policies, not Transformers models or a hosted inference endpoint. Elo calibration and provenance remain in the shipped board and MODELS.json.