File size: 7,271 Bytes
08646a9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 | {
"name": "recICL",
"model_type": "icl-recommender",
"architecture": "RecPFN",
"architecture_source": "https://github.com/SAP-samples/tabular-ai-recpfn (upstream commit 1af320c, Apache-2.0)",
"paper": "arXiv:2608.19735 (SIGIR '26), doi:10.1145/3805712.3809696",
"embedding_dim": 1536,
"n_parameters": 169918464,
"dtype": "float32",
"model_config": {
"n_layers": 4,
"n_heads": 8,
"dropout": 0.2,
"icl_module_type": "alternating",
"positional_embedding_scheme": "hard-alibi",
"max_sequence_len": 15,
"train_with_icl": true,
"num_icl_examples": 8,
"icl_k": 2,
"weighted_icl_sampling": true,
"weighted_decay_rate": 1.0,
"normalize_embs": true,
"llm": "qwen1b",
"baseline_config": "balanced"
},
"inference": {
"history_items_used": 14,
"context_sequence_max_len": 15,
"num_context_sequences": 8,
"context_retrieval": "SAP InvertedIndexMatcher: pool sequences cut into windows of 15 (step 7); candidates share one of the query's last 2 items; top 8 by recency-weighted Jaccard (decay 1.0)",
"score": "dot product of the output at the last history position with each L2-normalised item embedding"
},
"item_embeddings": {
"model_id": "Alibaba-NLP/gte-Qwen2-1.5B-instruct",
"revision": "a9af15a6372d7d6b25e9fb07c2ccb9e1fe645644",
"dim": 1536,
"trust_remote_code": true,
"compute_dtype": "bf16",
"max_seq_length": 512,
"prompt": null,
"pooling": "last token (<|endoftext|> appended by the remote-code tokenizer), right padding",
"normalization": "L2 in fp32",
"amazon_item_text_format": "Title: {title}; Brand: {brand}; Categories: {c1, c2, ...}"
},
"training": {
"code": "SAP-samples/tabular-ai-recpfn src/ at 1af320c + infrastructure patches (resume, logging, NaN diagnostics; opt-in fixes all OFF) - training math as released",
"config": {
"train_with_icl": true,
"num_icl_examples": 128,
"train_with_synthetic_data": true,
"n_layers": 4,
"batch_size": 16,
"reg_lambda": 0.0001,
"warmup_epochs": 6,
"baseline_config": "balanced",
"learning_rate": 0.0001,
"early_stopping_patience": 20,
"positional_embedding_scheme": "hard-alibi",
"optimizer": "adamw",
"llm": "qwen1b",
"max_sequence_len": 15,
"num_epochs": 120,
"use_cached_embeddings": true,
"num_gradient_accumulation_steps": 2,
"early_stopping_metric": "hr@10",
"loss": "sce",
"train_epoch_size": 500,
"val_epoch_size": 200,
"num_sdg_items": 1000,
"normalize_embs": true,
"n_heads": 8,
"dropout": 0.2,
"icl_module_type": "alternating",
"val_k": 10
},
"training_seed": 0,
"optimizer": "AdamW (torch defaults besides lr), linear warmup for warmup_epochs, then cosine annealing to 0.01*lr",
"data": "synthetic sequences only (no real user data); each batch = one synthetic environment of 1000 items whose embeddings are sampled from the embedding store below",
"embedding_store": {
"rows": 800000,
"composition": {
"fiqa": 64248,
"quora": 200000,
"fever": 200000,
"msmarco": 335752
},
"sources": "BEIR (FiQA-2018, Quora, FEVER, MS MARCO), sampled with seed 0",
"encoder": "same as item_embeddings",
"sha256": "fb9374462837c1b41188d15d50c55faf04096373f918ea31be57ddcdd6a29620"
},
"stages": {
"stage1": {
"synthetic_prior": "random graph",
"sdg_params": {
"transition_matrix_prior": [
"random_graph"
],
"popularity_bias": [
0.0,
0.1,
0.2,
0.3,
0.5
],
"window_size": [
1,
2,
3
],
"decay_coeff": [
0.5,
0.8
],
"repeated_items": [
true
]
},
"init": "random (seed 0)",
"epochs_run": 120,
"best_epoch": 101,
"early_stopped": false,
"best_synthetic_ratio_hr@10": 7.974883964010977,
"gpu": [
"NVIDIA H200"
],
"wall_hours": 2.573,
"first_epoch_end": "2026-09-29T07:33:10Z",
"finished": "2026-09-29T10:05:54Z"
},
"stage2": {
"synthetic_prior": "random graph + latent factor",
"sdg_params": {
"transition_matrix_prior": [
"random_graph",
"latent_factor"
],
"num_latent_concepts": [
20,
40,
60
],
"mean_num_concepts_per_item": [
2,
3,
5
],
"mean_num_concepts_per_user": [
2,
3,
5
],
"popularity_bias": [
0.0,
0.1,
0.2
],
"window_size": [
1,
2,
3
],
"decay_coeff": [
0.5,
0.8
],
"logit_scale": [
9.0,
12.0
],
"concept_matrix_diagonal_deviation_rate": [
0.0,
0.1,
0.3
],
"repeated_items": [
true
],
"mean_num_related_items": [
1,
3,
5,
7
],
"repeated_item_probability": [
0.1,
0.2,
0.3
]
},
"init": "stage-1 best checkpoint (epoch 101)",
"epochs_run": 47,
"best_epoch": 27,
"early_stopped": true,
"best_synthetic_ratio_hr@10": 3.9308144874246915,
"gpu": [
"NVIDIA H100 80GB HBM3"
],
"wall_hours": 1.099,
"first_epoch_end": "2026-09-29T10:07:51Z",
"finished": "2026-09-29T11:11:55Z"
}
},
"early_stopping": "hr@10 ratio (model / EmbKNN-balanced) on synthetic validation batches, patience 20",
"selection": {
"note": "best of our 9 training runs by mean validation HR@10 over the 7 Amazon validation splits",
"amazon7_val_hr@10": 0.13147873010138936,
"amazon7_val_hr@10_rank": 1,
"amazon7_val_mrr@10": 0.0884027748063644,
"amazon7_val_mrr@10_rank": 2,
"n_runs": 9
}
},
"evaluation_protocol": {
"split": "user-level 70/10/20, seed 42; at most 10,000 users sampled into the test split, then users with a single interaction dropped",
"target": "each test user's last item; input = up to 14 preceding items",
"context": "8 sequences from train-split users (retrieval as in 'inference')",
"batch_size": 1,
"ranking": "full item catalog, no candidate sampling, repeated items allowed",
"metrics": [
"HR@10",
"MRR@10"
],
"hardware": "NVIDIA L4, fp32, TF32 off"
},
"checkpoint": {
"source": "release_k4_s0/stage2/model.pth",
"source_sha256": "661fd3415a0b8f2d727898006e6314dcd273e15f59da3383668f7bc1a3997a3a",
"safetensors_sha256": "02c4edc3a4944a9e7a0b6025217b3ea7af16e53c9483e85e9a573074a228a60c",
"safetensors_bytes": 679680696,
"n_tensors": 64,
"src_fingerprint": "6c3ca6431e028e27"
}
}
|