File size: 6,046 Bytes
c7f732a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 | {
"name": "light-sufficiency-predictor-gguf8",
"status": "draft, not uploaded",
"exported": "2026-11-04",
"light_model": {
"base_model": "Qwen/Qwen3-0.6B (BF16)",
"license": "Apache-2.0",
"gguf_source": "unsloth/Qwen3-0.6B-GGUF/Qwen3-0.6B-BF16.gguf",
"backend": "llama",
"gguf_8_blocks": "model.gguf, the source GGUF cut to the first 8 blocks (token_embd, output_norm, output and blk.0 to blk.7)"
},
"language": "English only",
"layers": 8,
"llama_cpp_output": "per-token embeddings (pooling NONE), includes output_norm",
"prompt": {
"app_commit": "97272ef",
"language": "English prompt",
"target_words": 100,
"template": "<|im_start|>user\n<document>\n{text}\n</document>\n\nSummarize the document above in about 100 words (5 sentences). Use your own words and do not copy sentences from the document. Keep only the main points. Write only the summary.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n",
"document_placeholder": "{text}"
},
"tokenization": {
"add_bos": false,
"parse_special": true
},
"doc_span_rule": "Byte offsets of every token piece in the prompt (UTF-8, concatenated pieces equal the prompt bytes). The document text occupies bytes [len(prefix), len(prefix) + len(document bytes)) with prefix = '<|im_start|>user\\n<document>\\n'. A token is a document token when its byte range overlaps that range, so the token that merges the last document character with the following newline is included. The document text is used as given, not stripped.",
"pooling": "Mean over the document tokens of the per-token embeddings (llama.cpp pooling NONE, output after output_norm of block 8), one float32 vector of 1024 values.",
"probe": {
"file": "probe.json",
"formula": "z = w.x + b; p_sufficient = sigmoid(A * z + B); p_insufficient = 1 - p_sufficient",
"platt": {
"A": 0.9374034404754639,
"B": 0.22994965314865112
},
"fold_max_abs_diff": 3.743143861534293e-08
},
"label": "Label v2: sufficient (1) when the light summary passes the technical checks without too_long and the Gemma 3 12B judge scores faithfulness >= 4.",
"routing": {
"version": 1,
"points": [
[
0.0,
0.5
],
[
0.5,
0.8
],
[
1.0,
0.95
]
],
"interpolation": "linear",
"probability": "p_sufficient",
"rule": "light when p_sufficient >= threshold(quality_preference)",
"chosen_on": "validation (206 texts); Platt also fitted on validation, so optimistic"
},
"advice": {
"version": 1,
"mode": "none_with_prediction",
"rule": "with a prediction, do not suggest the heavy model; without one, the app's length advisor decides",
"why": "on validation the probe does not separate failures among light-routed texts (see threshold_analysis.md)"
},
"data": {
"train": {
"n": 616,
"insufficient": 104
},
"validation": {
"n": 206,
"insufficient": 30
},
"test": {
"n": 151,
"insufficient": 33
}
},
"validation_results": {
"n": 206,
"insufficient": 30,
"auroc": [
0.709659090909091,
0.5965633167023046,
0.8156790847945943
],
"ap_insufficient": [
0.40988197040299046,
0.25027936396129463,
0.5809268882725719
],
"hf_probe_auroc": [
0.7147727272727273,
0.6003268490904481,
0.8179235442920104
],
"note": "Platt is fitted on these validation texts; Brier and ECE are not reported on validation (in-sample)."
},
"test_results": {
"n": 151,
"insufficient": 33,
"models": {
"option 1 (GGUF 8 blocks)": {
"auroc": [
0.714175654853621,
0.6110126188571936,
0.8137687406296852
],
"ap_insufficient": [
0.4899080000407883,
0.3316064804585212,
0.6475444944318816
],
"brier": [
0.1548335387019752,
0.11751881178600608,
0.19731360723457353
],
"ece": [
0.05743244270615229,
0.035149464366451805,
0.1338328302005269
]
},
"HF probe": {
"auroc": [
0.7362609142270159,
0.6393254108153633,
0.8310708898944192
],
"ap_insufficient": [
0.5168968170940059,
0.354532133131363,
0.680021461893627
],
"brier": [
0.14991831587870777,
0.11258698218320425,
0.19120589099252627
],
"ece": [
0.06840796502220707,
0.03506714001988732,
0.13394637805539247
]
},
"(d)": {
"auroc": [
0.6612737544940934,
0.552839229640059,
0.7657977009611183
],
"ap_insufficient": [
0.4070139012278786,
0.2702898589196186,
0.5720459590368682
],
"brier": [
0.16557182463689082,
0.12485330876672007,
0.21032840020535495
],
"ece": [
0.059371011029426074,
0.03998848389548617,
0.14327356208017625
]
}
},
"diffs": {
"option 1 (GGUF 8 blocks) - (d)": [
0.052901900359527576,
-0.04098152794175529,
0.1478525632814101
],
"option 1 (GGUF 8 blocks) - HF probe": [
-0.022085259373394894,
-0.0400642421602785,
-0.004463789682539562
]
},
"reproduction": {
"HF probe": {
"max_abs_diff": 0.0,
"ok": true
},
"(d)": {
"max_abs_diff": 0.0,
"ok": true
}
},
"note": "Second use of the English test set, confirmation only."
},
"cost": {
"device": "llama.cpp, Metal (Apple GPU), median of 3 after a warmup",
"per_text": [
{
"input_tokens": 275,
"gguf8_embed_s": 0.072,
"app_full_summary_s": 0.97
},
{
"input_tokens": 941,
"gguf8_embed_s": 0.231,
"app_full_summary_s": 1.97
},
{
"input_tokens": 2494,
"gguf8_embed_s": 0.742,
"app_full_summary_s": 2.84
}
],
"note": "gguf8_embed_s covers prompt build, tokenize and the 8-block run; the probe itself is one dot product of 1024 values."
},
"parity": {
"texts": 20,
"split": "validation",
"domain": "government (GovReport)",
"max_abs_diff_probability": 4.1877598239148384e-07,
"mean_abs_diff_probability": 9.877418592485477e-08,
"tolerance": 0.0001,
"ok": true,
"span_texts_checked": 822,
"span_mismatches_vs_hf": 0,
"span_mismatch_ids": []
}
} |