{ "format": "streamtalk-retrained-cfg-v1", "architecture": "DiffusionDITNetPartsFixedExpressions2PostNormInteraction2", "parameter_count": 71167501, "state_dict": { "tensor_count": 407, "tensor_dtype": "torch.float32", "device": "cpu" }, "training": { "dataset": "BEAT2 English", "world_size": 4, "global_batch_size": 512, "cfg_dropout": 0.1, "training_code_commit": "b6cac99d5bae5164a21f5db25afef72c3c30556a" }, "evaluation": { "cfg_scale": 3.0, "speaker2_clip_count": 15, "all_clip_count": 265, "metric": "EMAGE/PantoMatrix AESK FGD", "combined_score_formula": "max(speaker2_fgd / 0.383, all_fgd / 0.293)", "protocol": "accelerated H200 evaluation; the final CFG=3 values have not been accepted against a direct paper-exact B=1/M=1/full-window oracle", "reporting_tolerance": "1e-3 absolute FGD (user-selected engineering tolerance; not a completed oracle validation)", "published_inference_data_rescore": { "metric_batch_size": 16, "speaker2_fgd": 0.3788789702354345, "speaker2_selection_absolute_difference": 7.105427357601002e-15, "all_fgd": 0.21764184426140076, "all_selection_absolute_difference": 0.00002966349321199 }, "seed": 0 }, "artifacts": [ { "file": "streamtalk_speaker2_combined_e0946_cfg3.pt", "roles": ["speaker2", "combined"], "completed_epoch": 946, "cfg_scale": 3.0, "speaker2_fgd": 0.3788789702354416, "all_fgd": 0.2502867250611853, "combined_score": 0.9892401311630329, "size_bytes": 284796631, "sha256": "07065d52e333cc48960e4e8fad5c0331f6b9e482f14aa1b78cb721904da5b075" }, { "file": "streamtalk_speaker_all_e0940_cfg3.pt", "roles": ["speaker_all"], "completed_epoch": 940, "cfg_scale": 3.0, "speaker2_fgd": 0.4244774977704626, "all_fgd": 0.21767150775461275, "combined_score": 1.108296338826273, "size_bytes": 284796631, "sha256": "3a0e3558d788fd3f11a5a87a8198b42725104b5b9b24f95a4e7c84c14395c9a5" } ] }