{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreidhtpjnpjjg3725z5qf36e4qtj2e2xfe7egnshivr266vbrfr23ae",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mop7vd3xm632"
},
"path": "/t/datasets-and-the-right-models/176969#post_4",
"publishedAt": "2026-06-20T06:20:18.000Z",
"site": "https://discuss.huggingface.co",
"textContent": "I’d try Llama 3.1/3.2 8B Instruct or Mistral 7B. They generally handle structured classification and abstention better than generation-focused tuning. Also, include plenty of “no-match” examples so the model learns when to abstain instead of always picking a candidate.",
"title": "Datasets and the right models"
}