alexwengg's picture
short-reply-0.6b Core ML package (prefill on ANE, decode GPU), tokenizer, host config
db47ae8 verified
Raw History Blame Contribute Delete
523 Bytes
{
"package": "short_reply_0_6b.mlpackage",
"prefillLength": 160,
"cacheLength": 512,
"decodeLengths": [1, 64, 128, 256],
"layers": 28,
"kvHeads": 8,
"headDim": 128,
"padID": 151643,
"stopIDs": [151645, 151643],
"maxNewTokens": 32,
"systemPrompt": "Write one natural, very short reply to the social post as a separate person. React to a concrete detail in the post when possible. Do not invent facts or personal experiences. Use at most 12 words. Output only the reply.",
"maxPostCharacters": 2000
}