Experimental browser for the Atmosphere
If you trained your tokenizer on Reddit data, but not your LLM then the problem could persist. It seems like the deletion step should be a solid win, except that the inference procedure changes a bit, so the current fast HuggingFace tokenizers can't be used.
Dec 20, 2024, 6:45 PM
{
"text": "If you trained your tokenizer on Reddit data, but not your LLM then the problem could persist. \n\nIt seems like the deletion step should be a solid win, except that the inference procedure changes a bit, so the current fast HuggingFace tokenizers can't be used.",
"$type": "app.bsky.feed.post",
"langs": [
"en"
],
"reply": {
"root": {
"cid": "bafyreictfbybzbhtvpo4cfus4h45zrh64b3qqgcnfr5phmn3wimmbrhgdm",
"uri": "at://did:plc:54lqvssae6v2kio2cu26yktz/app.bsky.feed.post/3ldr2b5ec722r"
},
"parent": {
"cid": "bafyreidys4w2xn4l5rxyrghph6upr72s7j62um64mz2ur5e2apm3sgeh6m",
"uri": "at://did:plc:54lqvssae6v2kio2cu26yktz/app.bsky.feed.post/3ldr2b5scgs2r"
}
},
"createdAt": "2024-12-20T18:45:49.351Z"
}