ut-depth-probe-artifacts / loopq_quantization /scripts /configs /paper_ouro_reference_metrics.json
JunYoungLee's picture
Add LoopQ 4-bit quantization of Ouro-1.4B
9118991 verified
Raw History Blame Contribute Delete
1.7 kB
{
"source": "arXiv:2605.16343v1 Table 1",
"model": "Ouro 1.4B",
"source_tex_sha256": "3e6277435595b30e8f812b00c776dec5be3d45611088af0093f64c4b60be1f30",
"accuracy_protocol": "Paper says acc_norm when available, otherwise acc; ten-run accuracy averages",
"rows": {
"BF16": {
"W4A4": {
"HS": 0.7159,
"WG": 0.6717,
"LB": 0.6505,
"AC": 0.5256,
"MM": 0.6823,
"WT2": 12.03,
"OP": 5.3
},
"W4A8": {
"HS": 0.7159,
"WG": 0.6717,
"LB": 0.6505,
"AC": 0.5256,
"MM": 0.6823,
"WT2": 12.03,
"OP": 5.3
}
},
"Symmetric": {
"W4A4": {
"HS": 0.2777,
"WG": 0.4941,
"LB": 0.0033,
"AC": 0.2389,
"MM": 0.2282,
"WT2": 1170.0,
"OP": 82000.0
},
"W4A8": {
"HS": 0.5022,
"WG": 0.5866,
"LB": 0.322,
"AC": 0.3745,
"MM": 0.3569,
"WT2": 43.89,
"OP": 46.72
}
},
"LoopQ": {
"W4A4": {
"HS": 0.6582,
"WG": 0.5951,
"LB": 0.5659,
"AC": 0.4881,
"MM": 0.5465,
"WT2": 15.81,
"OP": 8.18
},
"W4A8": {
"HS": 0.6954,
"WG": 0.6456,
"LB": 0.6255,
"AC": 0.5145,
"MM": 0.6206,
"WT2": 13.29,
"OP": 6.09
}
}
},
"local_comparison_caveats": [
"Local BF16 runs use explicit single-seed protocols.",
"Local direct RTN uses dynamic per-input absmax activation scales; do not equate it with the paper Symmetric row.",
"WikiText perplexity definitions and tokenization require explicit matching."
]
}