From 12a391d324e2a37470233df8e6aeef038e680cbb Mon Sep 17 00:00:00 2001
From: "Shiyi Zheng (from Dev Box)" <shzhen@microsoft.com>
Date: Tue, 23 Jun 2026 10:57:32 +0800
Subject: [PATCH] examples: add facebook/bart-large-mnli text-classification
 recipe

Ships an fp32 NLI head for facebook/bart-large-mnli at task=text-classification.
Recipe carries value_range=[2,3] on input_ids to deterministically inject the
eos_token_id required by BartForSequenceClassification's eos-pooling head.

Goal-ladder verdict (CPU):
- L0 build PASS  - 1042 ops, 21 unique types, 407M params, 384 KB graph + 1.6 GB external data
- L1-CPU perf PASS - 1.64 s/iter on 1024-token real-tokenized input
                     (custom Python script; winml perf ignores recipe value_range
                     and crashes on eos-pooling models with random ints - winml
                     CLI feature gap to file separately)
- L2 numerical PASS - cosine = 1.000000, max_abs = 1e-6 vs PyTorch reference
                      (argmax = 2, ENTAILMENT, on both sides)
- L3 task-metric PASS - accuracy = 0.88, latency = 1.89 s/sample on
                        glue/mnli/validation_matched/100 samples, seed=42
                        (matches published ~0.886 within MC noise; first end-to-end
                        Goal-L3 PASS for this repo)

DML/QNN/OpenVINO are HOST-BLOCKED on producer host (DML 0xC0000409, QNN absent,
OpenVINO DLL-load-fails) - not penalized per local skill convention.

Optimum-coverage: VENDOR-COVERED on text-classification via Optimum BartOnnxConfig;
recipe is pure-data, no per-architecture code change needed.

Producer notes from running the recipe live in research/adding-model-support/
model_knowledge/bart.json on the skills-poc working branch (not landed to main
yet; pending separate skill-research PR for the full research/ tree).
---
 examples/recipes/README.md                    |  1 +
 .../text-classification_config.json           | 58 +++++++++++++++++++
 2 files changed, 59 insertions(+)
 create mode 100644 examples/recipes/facebook_bart-large-mnli/text-classification_config.json

diff --git a/examples/recipes/README.md b/examples/recipes/README.md
index caaa2c15f..ead4e01a2 100644
--- a/examples/recipes/README.md
+++ b/examples/recipes/README.md
@@ -17,6 +17,7 @@ Each *(model, task)* includes:
 | Model | Task |
 |---|---|
 | BAAI/bge-large-en-v1.5 | sentence-similarity |
+| facebook/bart-large-mnli | text-classification |
 | cardiffnlp/twitter-roberta-base-sentiment-latest | text-classification |
 | deepset/roberta-base-squad2 | question-answering |
 | deepset/tinyroberta-squad2 | question-answering |
diff --git a/examples/recipes/facebook_bart-large-mnli/text-classification_config.json b/examples/recipes/facebook_bart-large-mnli/text-classification_config.json
new file mode 100644
index 000000000..7d2395177
--- /dev/null
+++ b/examples/recipes/facebook_bart-large-mnli/text-classification_config.json
@@ -0,0 +1,58 @@
+{
+  "_note": "Workaround: input_ids.value_range narrowed to [2,3] (eos_token_id=2) so BartForSequenceClassification eos-pooling never sees an empty match during export. Verified end-to-end: build complete (157s), L0 PASS (1042 nodes, opset 17), L1-CPU PASS (1638ms/iter with real input — `winml perf` will fail because random ints lack eos), L2 PASS (cosine=1.0, max_abs=1e-6 vs PyTorch on premise+hypothesis pair). See research/adding-model-support/model_knowledge/bart.json bart-004.",
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "input_ids",
+        "dtype": "int32",
+        "shape": [
+          1,
+          1024
+        ],
+        "value_range": [
+          2,
+          3
+        ]
+      },
+      {
+        "name": "attention_mask",
+        "dtype": "int32",
+        "shape": [
+          1,
+          1024
+        ],
+        "value_range": [
+          1,
+          2
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "logits"
+      }
+    ]
+  },
+  "optim": {
+    "clamp_constant_values": true,
+    "gelu_fusion": true,
+    "matmul_add_fusion": true,
+    "remove_isnan_in_attention_mask": true
+  },
+  "quant": null,
+  "compile": null,
+  "loader": {
+    "task": "text-classification",
+    "model_class": "AutoModelForSequenceClassification",
+    "model_type": "bart"
+  }
+}