clips / clips_meta.json
bkicanao's picture pveugen's picture
Update repository
a3ca71d
Raw History Blame
2.92 kB
{
"version": "2.0.0",
"task": "clip-selection",
"trunk": "xlm-roberta-base",
"checkpoint_digest": "fe11c7852ef0421e9a0386822e32a9cbab43b422bc3ac4759dba42f5fda50d5b",
"quantization": {
"note": "The two platforms use DIFFERENT int8 schemes. A single `quantization` field here said `int8_linear_perchannel` and covered both, which was wrong for the LiteRT files.",
"core_ml": "int8_linear_perchannel",
"litert": "weight_only_wi8_afp32"
},
"evaluation": {
"core_ml": "475 clips judged on 20 transcripts from the frozen holdout, one judge family (Gemma). No human read.",
"litert": "NONE. No clip has been read from the LiteRT artifacts on any platform. Their only gate is a synthetic random-token batch, whose own manifest records `is_a_quality_result: false`."
},
"artifacts": {
"core_ml": {
"file": "clips.mlmodelc",
"note": "ONE multifunction package. Both graphs share one stored copy of the trunk, so reaching a graph needs MLModelConfiguration.functionName - a path alone names the file, not the graph, and Core ML will silently load the default function.",
"functions": {
"select": {
"inputs": {"ids": [16, 128], "mask": [16, 128], "disc": [16, 5]},
"outputs": ["saliency", "start_p", "end_p"]
},
"score": {
"inputs": {"ids": [16, 256], "mask": [16, 256]},
"outputs": ["score"]
}
}
},
"litert": {
"note": "LiteRT has no multifunction packaging, so these platforms ship two files and store the shared trunk twice.",
"selector": {
"file": "clips-selector.tflite",
"signature": "serving_default",
"inputs": {"args_0": [16, 128], "args_1": [16, 128], "args_2": [16, 5]},
"outputs": ["saliency", "start_p", "end_p"]
},
"scorer": {
"file": "clips-scorer.tflite",
"signature": "serving_default",
"inputs": {"args_0": [16, 256], "args_1": [16, 256]},
"outputs": ["score"]
}
}
},
"disc_features": [
"relative_position",
"hook_cue_count",
"payoff_cue_count",
"ends_with_question",
"digit_ratio"
],
"tokenizer": {
"file": "clip_tokenizer.bin",
"format": "RDTK",
"sha256": "f14166c0b746be5e4581954bfbbf0ec04b54c1eae3485f7abd8f62aa28a38fb4",
"vocab_size": 250002,
"unk_id": 3,
"bos_id": 0,
"eos_id": 2,
"pad_id": 1
},
"windows": {
"selector_seq_len": 128,
"scorer_seq_len": 256,
"sentence_truncation": 64,
"note": "sentence_truncation is the token cap applied to ONE sentence and is NOT the graph width. The saliency heads were trained at 64; feeding a longer single sentence runs them off-distribution."
},
"recommended": {
"min_sentences_per_clip": 2,
"words_per_second_proxy": 2.5,
"default_clip_limit": 10,
"clip_limit_is": "a cap, not a quota - shorter videos return fewer"
}
}