| { |
| "model_name": "Faynt-10M-Arena", |
| "source_report": "Faynt: Scaling and Optimizing Policies for Competitive Melee", |
| "initial_suite": { |
| "games_per_policy": 152, |
| "results": [ |
| { |
| "model": "Faynt-10M-Base", |
| "wins": 49, |
| "test_controller_nll_nats_per_target": 0.79766, |
| "winner_weighted_validation_nll": 0.81152 |
| }, |
| { |
| "model": "Faynt-75M-Base", |
| "wins": 69, |
| "test_controller_nll_nats_per_target": 0.76543, |
| "winner_weighted_validation_nll": 0.77502 |
| }, |
| { |
| "model": "Faynt-10M-Expert", |
| "wins": 106, |
| "test_controller_nll_nats_per_target": 0.76974, |
| "winner_weighted_validation_nll": 0.75746 |
| }, |
| { |
| "model": "Faynt-75M-Expert", |
| "wins": 123, |
| "test_controller_nll_nats_per_target": 0.74206, |
| "winner_weighted_validation_nll": 0.72489 |
| } |
| ], |
| "source": "Section 5.1, Initial Benchmark Suite, Figure 8; Appendix F.2, The Original 152-Game Panel, F.3, Execution, Scoring, and Reuse, and F.7, Fixed-Panel Win and Stock Ledgers; Appendix E.4, Selected Checkpoints and Slice Diagnostics, Tables E.7 and E.8", |
| "sampling_seeds_shared": true, |
| "emulator_starting_rng_fully_controlled": false, |
| "fidelity_audit_status": "complete", |
| "verified_outcomes_per_policy": 152 |
| }, |
| "expanded_suite": { |
| "games_per_arena_policy": 1312, |
| "opponent_releases": 14, |
| "conditions": { |
| "supported_mirrors": { |
| "games": 244, |
| "10M_wins": 240, |
| "75M_wins": 149 |
| }, |
| "extended_roster": { |
| "games": 534, |
| "10M_wins": 427, |
| "75M_wins": 198 |
| }, |
| "forced_mirrors": { |
| "games": 534, |
| "10M_wins": 531, |
| "75M_wins": 503 |
| } |
| }, |
| "source": "Sections 5.2, Final Benchmark After Reinforcement Learning, and 5.3, Supported-Character and Full-Roster Performance, Figure 9; Appendix F.5, Expanded Slippi-AI Panel, Table F.5, F.8, Supported-Character Mirrors, F.9, Extended-Roster Matches, and F.10, Forced-Character Mirrors", |
| "faynt_added_delay_frames": 0, |
| "opponent_delay_frames": [ |
| 21, |
| 24 |
| ], |
| "delay_effect_isolated": false |
| }, |
| "zero_delay": { |
| "private_slippi_ai": { |
| "games_per_policy": 68, |
| "10M_Expert_wins": 23, |
| "75M_Expert_wins": 25, |
| "10M_Arena_wins": 68, |
| "75M_Arena_wins": 58, |
| "10M_Arena_four_stock_wins": 61, |
| "source": "Sections 5.4, Zero-Delay Comparisons, and 6.5, From Supervised Post-Training to Reinforcement Learning, Figure 15; Appendix F.13, Privately Supplied Zero-Delay Slippi-AI Checkpoint, Table F.8" |
| }, |
| "phillip_final_destination": { |
| "games": 112, |
| "10M_wins": 112, |
| "75M_wins": 108 |
| }, |
| "phillip_six_stages": { |
| "games": 126, |
| "10M_wins": 126, |
| "75M_wins": 126 |
| }, |
| "source": "Section 5.4, Zero-Delay Comparisons; Appendix F.12, Zero-Delay Specialist Evaluation, Table F.7, and F.13, Privately Supplied Zero-Delay Slippi-AI Checkpoint, Table F.8", |
| "phillip_action_repeat_frames": [ |
| 2, |
| 3, |
| 4 |
| ], |
| "faynt_action_repeat_frames": 1 |
| }, |
| "limitations": [ |
| "Base and Expert use the initial suite; Arena uses the expanded suite.", |
| "Model sizes have different training histories.", |
| "Some opponents informed RL selection.", |
| "Forced mirrors use characters outside the opponent deployed roster.", |
| "Additional-character expanded cells outside the opponent deployed roster have two games; supported specialist mirrors have 12 games each." |
| ], |
| "behavior_analysis": { |
| "cohort_games": { |
| "10M": 139, |
| "75M": 145 |
| }, |
| "scope": "Frozen matched Base-to-Expert analysis cohorts; all152initial-panel outcomes are verified.", |
| "source": "Appendix G.1, Replay Selection and Evaluation Conditions; Tables G.1-G.4" |
| }, |
| "benchmark_actor": { |
| "context_frames": 256, |
| "context_mode": "ring", |
| "added_policy_delay_frames": 0, |
| "action_sampling_temperature": 1.0, |
| "cache_behavior": "Continuous native key/value ring; capacity comes from model_config.context_length." |
| }, |
| "actor_trajectory_context": { |
| "frames": 128, |
| "meaning": "Actor trajectory/history configuration, separate from the 256-frame native ring-cache capacity." |
| }, |
| "checkpoint_source_sha256": "c8a810c0d33b72ae448eec0e0b7d0538beacfaeb9a2e357eddd7be4fc0ce286b" |
| } |
|
|
