Release v1.3.0: FastDecider-149M weights, tokenizer, and model card

README.md

@@ -0,0 +1,98 @@
+---
+language:
+- en
+- zh
+tags:
+- modernbert
+- reranker
+- cross-encoder
+- agent
+- decision-making
+- zero-shot-classification
+- fast-decider
+license: apache-2.0
+pipeline_tag: text-classification
+library_name: transformers
+metrics:
+- accuracy
+- brier_score
+model-index:
+- name: FastDecider-149M
+  results:
+  - task:
+      type: text-classification
+      name: Agent Decision Ranking
+    dataset:
+      name: JevBench v1.2.11
+      type: jevbench
+    metrics:
+    - type: accuracy
+      value: 100.0
+      name: Easy Tier Accuracy
+    - type: accuracy
+      value: 44.14
+      name: Hard Tier Accuracy
+    - type: accuracy
+      value: 68.06
+      name: Original Tier Accuracy
+---
+
+# FastDecider-149M (v1.3.0): Sub-15ms System-1 Neural Decision Engine
+
+**FastDecider-149M** is a high-speed (14.93ms P50 latency), 149M-parameter neural Cross-Encoder designed for autonomous agents, browser DOM interaction, API routing, and high-throughput operational decisions.
+
+## Benchmark Performance
+
+- **JevBench Easy Tier**: **100.0%** (48/48)
+- **JevBench Hard Tier**: **44.14%** (49/111)
+- **JevBench Original Tier**: **68.06%** (49/72)
+- **JevBench Composite Score**: **78.15 / 100** (Rank #1 Composite)
+- **P50 Latency**: **14.93 ms** on GPU
+- **Cost**: **$0.0012** per 1,000 decisions
+- **6 Core Business Pillars**:
+  - Browser DOM Control: 100.0% (50/50)
+  - API Tool Dispatch: 98.0% (49/50)
+  - E-Commerce Brands: 100.0% (50/50)
+  - E-Commerce Specs: 100.0% (50/50)
+  - E-Commerce Category: 94.0% (47/50)
+  - Legal Contracts: 96.0% (48/50)
+
+## Usage with Transformers
+
+```python
+import torch
+import torch.nn.functional as F
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
+model_id = "mkzero/FastDecider-149M"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForSequenceClassification.from_pretrained(
+    model_id, dtype=torch.bfloat16
+).to("cuda")
+model.eval()
+
+context = "User wishes to cancel order #4821 within 2 hours of checkout."
+instruction = "Select the appropriate API handler."
+options = {
+    "cancel_order": "orders.cancel_immediate",
+    "request_return": "returns.create_ticket",
+    "contact_support": "support.live_agent"
+}
+
+pairs = [(f"Context:
+{context}
+
+Instruction:
+{instruction}", f"Option {k}: {v}") for k, v in options.items()]
+inputs = tokenizer(pairs, padding=True, truncation=True, max_length=2048, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    logits = model(**inputs).logits.squeeze(-1).float()
+    probs = F.softmax(logits, dim=-1).cpu().numpy()
+
+keys = list(options.keys())
+print("Decision:", keys[probs.argmax()])
+```
+
+## GitHub Repository
+Source code, benchmarks, and microservice serving: [https://github.com/mkzero/FastDecider-149M](https://github.com/mkzero/FastDecider-149M)

config.json

@@ -0,0 +1,83 @@
+{
+  "architectures": [
+    "ModernBertForSequenceClassification"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 50281,
+  "classifier_activation": "gelu",
+  "classifier_bias": false,
+  "classifier_dropout": 0.0,
+  "classifier_pooling": "mean",
+  "cls_token_id": 50281,
+  "decoder_bias": true,
+  "deterministic_flash_attn": false,
+  "dtype": "bfloat16",
+  "embedding_dropout": 0.0,
+  "eos_token_id": 50282,
+  "global_attn_every_n_layers": 3,
+  "gradient_checkpointing": false,
+  "hidden_activation": "gelu",
+  "hidden_size": 768,
+  "id2label": {
+    "0": "LABEL_0"
+  },
+  "initializer_cutoff_factor": 2.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 1152,
+  "label2id": {
+    "LABEL_0": 0
+  },
+  "layer_norm_eps": 1e-05,
+  "layer_types": [
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "local_attention": 128,
+  "max_position_embeddings": 8192,
+  "mlp_bias": false,
+  "mlp_dropout": 0.0,
+  "model_type": "modernbert",
+  "norm_bias": false,
+  "norm_eps": 1e-05,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 22,
+  "pad_token_id": 50283,
+  "position_embedding_type": "absolute",
+  "rope_parameters": {
+    "full_attention": {
+      "rope_theta": 160000.0,
+      "rope_type": "default"
+    },
+    "sliding_attention": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    }
+  },
+  "sep_token_id": 50282,
+  "sparse_pred_ignore_index": -100,
+  "sparse_prediction": false,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.15.1",
+  "vocab_size": 50368
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff1ecc3791ada36b20c904710696885cd1b4200e931f9b8a62db5c1221f4e40e
+size 299225554

model_card.json

@@ -0,0 +1,62 @@
+{
+  "model_name": "FastDecider-149M",
+  "version": "1.3.0",
+  "archive_date": "2026-09-22",
+  "base_model": "Alibaba-DAMO/gte-reranker-modernbert-base",
+  "base_model_architecture": "ModernBERT-base Cross-Encoder (SequenceClassification)",
+  "parameter_count": 149079553,
+  "parameter_size_mb": "149M",
+  "num_layers": 22,
+  "hidden_size": 768,
+  "max_position_embeddings": 8192,
+  "training_details": {
+    "dataset": "/mnt/workspace/data/grand_unified_train_35k.jsonl",
+    "total_samples": 30456,
+    "trainer_script": "/mnt/workspace/arena/train_grand_unified_35k.py",
+    "training_steps": 600,
+    "optimizer": "AdamW (lr=2e-5, warmup=50, cosine schedule)",
+    "loss_function": "Cross-Entropy over candidate option logits",
+    "precision": "bfloat16"
+  },
+  "benchmarks": {
+    "jevbench_public_easy": {
+      "accuracy": "48/48 (100.00%)",
+      "extraction": "12/12 (100.0%)",
+      "fact": "12/12 (100.0%)",
+      "intent": "12/12 (100.0%)",
+      "tool_selection": "12/12 (100.0%)"
+    },
+    "jevbench_public_hard": {
+      "accuracy": "49/111 (44.14%)",
+      "routing_hard": "5/5 (100.0%)",
+      "tradeoff": "6/6 (100.0%)",
+      "trap": "5/8 (62.5%)",
+      "ambiguous": "4/7 (57.1%)",
+      "adversarial": "3/6 (50.0%)",
+      "probability": "5/10 (50.0%)",
+      "judge_hard": "7/17 (41.2%)",
+      "temporal_numeric": "5/15 (33.3%)",
+      "long_policy": "6/19 (31.6%)",
+      "multi_hop": "3/18 (16.7%)"
+    },
+    "jevbench_public_original": {
+      "accuracy": "49/72 (68.06%)",
+      "intelligence_score": 65.35
+    },
+    "six_business_pillars": {
+      "browser_dom_control": "50/50 (100.0%)",
+      "api_dispatch_routing": "49/50 (98.0%)",
+      "ecom_brand_extraction": "50/50 (100.0%)",
+      "ecom_specs_matching": "50/50 (100.0%)",
+      "ecom_category_prediction": "47/50 (94.0%)",
+      "legal_contract_matching": "48/50 (96.0%)"
+    },
+    "latency_and_cost": {
+      "p50_latency_ms": 14.93,
+      "p99_latency_ms": 17.8,
+      "serving_cost": "$0.0012 per 1k decisions",
+      "throughput_qps": 84.2
+    }
+  },
+  "reproducibility": "All scores evaluated via standard argmax over cross-encoder logits without heuristic overrides."
+}

tokenizer_config.json

@@ -0,0 +1,24 @@
+{
+  "backend": "tokenizers",
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "is_local": true,
+  "local_files_only": false,
+  "mask_token": "[MASK]",
+  "max_length": 512,
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 8192,
+  "pad_to_multiple_of": null,
+  "pad_token": "[PAD]",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "sep_token": "[SEP]",
+  "stride": 0,
+  "tokenizer_class": "TokenizersBackend",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "[UNK]"
+}