Publish the TOPv2-tuned MiniLM encoder for Ren

1_Pooling/config.json

@@ -0,0 +1,5 @@
+{
+    "embedding_dimension": 384,
+    "pooling_mode": "mean",
+    "include_prompt": true
+}

README.md

@@ -0,0 +1,65 @@
+---
+license: cc-by-sa-4.0
+language:
+- en
+library_name: sentence-transformers
+pipeline_tag: sentence-similarity
+base_model: sentence-transformers/all-MiniLM-L6-v2
+tags:
+- sentence-transformers
+- intent-classification
+- calibration
+- selective-prediction
+---
+
+# Secnd Ren encoder (MiniLM tuned on TOPv2)
+
+The sentence encoder behind **Ren** (`secnd-ren-1`), Secnd's decision model. Ren learns one organisation's own past decisions with a small logistic-regression head on top of this encoder, and answers on its own only where a calibrated gate can prove 99% precision; everything else goes to a person.
+
+This encoder is [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) tuned once on Meta's TOPv2 assistant requests with a supervised contrastive loss. It has the same size and CPU cost as the base model (384 dimensions, 22.7M parameters).
+
+## Training
+
+- **Data:** TOPv2 (Chen et al., 2020), English task-oriented requests in 8 domains, CC BY-SA 4.0. Label = the top-level intent of each parse (69 intents).
+- **No MTOP rows seen:** every TOPv2 request whose normalised text appears anywhere in MTOP was removed first (23,137 rows), leaving 157,405. MTOP results below therefore use only the labels inside each budget.
+- **Objective:** supervised contrastive loss over class-balanced batches (16 intents x 4 rows), temperature 0.05, AdamW lr 2e-5, 3,000 steps, seed 0, one NVIDIA A10G.
+
+## Results (MTOP English, held-out test)
+
+Ren with this encoder, 2,000 training labels, fixed calibration and test splits:
+
+| Encoder | Answered automatically | Precision there | Accuracy | 99% claim holds across 2,000 re-splits |
+|---|---|---|---|---|
+| all-MiniLM-L6-v2 | 31.7% | 99.62% | 86.2% | 99.05% |
+| **this encoder** | **42.4%** | **99.93%** | **88.1%** | **99.45%** |
+
+In a simulated learning loop on MTOP (200 seed labels, then 20 rounds of 100 answers, 5 seeds), the share answered automatically at round 20 rose from 19.2% to 33.9%, with test precision at least 99.24% in every round that answered automatically.
+
+Measurements and the code that produced them: [github.com/Rish-it/secnd](https://github.com/Rish-it/secnd) (`docs/results/s3/card-mtop-pt-minilm+lr-b2000.json`, `docs/results/s3/robustness-mtop-pt-minilm+lr-b2000.json`, `docs/results/s4/`).
+
+## Use
+
+```python
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer("Rishh00/secnd-ren-encoder")
+vectors = model.encode(["set an alarm for 7am", "what's the weather tomorrow"], normalize_embeddings=True)
+```
+
+## Limits
+
+- English only. Tuned on assistant-style requests; other domains may gain less (or nothing).
+- The 99% figure is a property of Ren's calibrated gate on a fixed, published split, not of this encoder alone, and not a guarantee on arbitrary future traffic.
+
+## Licence and citation
+
+Released under CC BY-SA 4.0 because it is trained on TOPv2 (CC BY-SA 4.0). The base model is Apache-2.0.
+
+```bibtex
+@inproceedings{chen-etal-2020-low,
+  title = {Low-Resource Domain Adaptation for Compositional Task-Oriented Semantic Parsing},
+  author = {Chen, Xilun and Ghoshal, Asish and Mehdad, Yashar and Zettlemoyer, Luke and Gupta, Sonal},
+  booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
+  year = {2020}
+}
+```

config.json

@@ -0,0 +1,30 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "BertModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "is_decoder": false,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 6,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.21.0",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522
+}

config_sentence_transformers.json

@@ -0,0 +1,14 @@
+{
+  "__version__": {
+    "pytorch": "2.15.0+cu132",
+    "sentence_transformers": "5.7.0",
+    "transformers": "5.21.0"
+  },
+  "default_prompt_name": null,
+  "model_type": "SentenceTransformer",
+  "prompts": {
+    "document": "",
+    "query": ""
+  },
+  "similarity_fn_name": "cosine"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f76f276dc59040cf7c3fe1ac74d884b70fed3c60fe2eaa3738adb1f7ae35c089
+size 90864192

modules.json

@@ -0,0 +1,20 @@
+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "sentence_transformers.base.modules.transformer.Transformer"
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Pooling",
+    "type": "sentence_transformers.sentence_transformer.modules.pooling.Pooling"
+  },
+  {
+    "idx": 2,
+    "name": "2",
+    "path": "2_Normalize",
+    "type": "sentence_transformers.sentence_transformer.modules.normalize.Normalize"
+  }
+]

sentence_bert_config.json

@@ -0,0 +1,10 @@
+{
+    "transformer_task": "feature-extraction",
+    "modality_config": {
+        "text": {
+            "method": "forward",
+            "method_output_name": "last_hidden_state"
+        }
+    },
+    "module_output_name": "token_embeddings"
+}

tokenizer_config.json

@@ -0,0 +1,24 @@
+{
+  "backend": "tokenizers",
+  "cls_token": "[CLS]",
+  "do_basic_tokenize": true,
+  "do_lower_case": true,
+  "is_local": false,
+  "local_files_only": false,
+  "mask_token": "[MASK]",
+  "max_length": 128,
+  "model_max_length": 256,
+  "never_split": null,
+  "pad_to_multiple_of": null,
+  "pad_token": "[PAD]",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "sep_token": "[SEP]",
+  "stride": 0,
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "[UNK]"
+}