From 5e6eeda058395c08bc4cce8beda04b3cec2c6f83 Mon Sep 17 00:00:00 2001 From: Niroj33 <86530493+professor3333@users.noreply.github.com> Date: Sun, 15 Nov 2026 08:32:55 +0545 Subject: [PATCH] Link the downloadable model bundles from the README and model card The fine-tuned run-3 DeBERTa-v3-small is published as the model-only pre-release model-deberta-3-run3 (not the default; checksums recorded). The model card gains a Downloads section with both bundles and corrects the CRF row, which still described the pilot-only model. --- README.md | 7 +++++++ docs/model_card.md | 15 ++++++++++++++- 2 files changed, 21 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 4005c04..60e1410 100644 --- a/README.md +++ b/README.md @@ -408,6 +408,13 @@ acceptance gate has been evaluated, because labels after the pilot are model dra [release notes](docs/release_notes/v0.1.0.md). Its rollback drill (rules `r1` → CRF `v0.1.0` → `r1`) passed on local Docker. +**Model downloads.** The CRF bundle is attached to +[v0.1.0](https://github.com/professor3333/TriggerMap/releases/tag/v0.1.0). The fine-tuned run-3 +DeBERTa-v3-small (weights, tokenizer, labels, calibration, provenance, MIT notice) is a separate +model-only pre-release, +[`model-deberta-3-run3`](https://github.com/professor3333/TriggerMap/releases/tag/model-deberta-3-run3). +It is not the default model. Checksums are in [`docs/model_card.md`](docs/model_card.md#downloads). + ## Serving cost **On the target host (2 vCPU / 4 GB x86-64, AWS EC2, 2026-11-09)**, the released CRF diff --git a/docs/model_card.md b/docs/model_card.md index d29737a..a47f91c 100644 --- a/docs/model_card.md +++ b/docs/model_card.md @@ -28,7 +28,7 @@ reviewer confirms or corrects every item before export. | System | What it is | Status | |---|---|---| | `rules-2` | Deterministic cue rules over the shared preprocessing (no learning) | Runs; 100% formatting-consistent on 300 documents | -| CRF | `python-crfsuite` BIO span tagger with logistic-regression category, role and group classifiers. Features `crf-features-1`, optional spaCy part-of-speech tags | Trained on pilot data only | +| CRF | `python-crfsuite` BIO span tagger with logistic-regression category, role and group classifiers. Features `crf-features-1`, optional spaCy part-of-speech tags | Default model: `crf-batches-01-02` (train split of `pilot-2`, `batch-01`, `batch-02`; 239 records), released in v0.1.0 | | DistilBERT | `distilbert/distilbert-base-cased` (Apache-2.0), revision `6ea81172465e8b0ad3fddeed32b986cdcdcffcf0` | Run 3: held-out agreement typed F1 0.099 (CRF 0.127), [`reports/neural/run-3/`](../reports/neural/run-3/README.md) | | DeBERTa-v3-small | `microsoft/deberta-v3-small` (MIT), revision `a36c739020e01763fe789b4b85e2df55d6180012` | Run 3: held-out agreement typed F1 0.185, false-prerequisite rate 0.483 (CRF 0.127 / 0.412; differences not clear), [`reports/neural/run-3/`](../reports/neural/run-3/README.md) | @@ -43,6 +43,19 @@ The transformer models share one design: The configs are in `configs/experiments/`. All systems end in the same deterministic validator, which rejects invalid offsets, dangling group members and contradictory groups. +## Downloads + +| Bundle | Release | File | SHA-256 | +|---|---|---|---| +| CRF `crf-batches-01-02` (default) | [v0.1.0](https://github.com/professor3333/TriggerMap/releases/tag/v0.1.0) | `crf-batches-01-02.tgz` (861,275 bytes) | `65677a7413e02a8da4ba3b920ac7eb8e02a8cba874876d93810d4ae5fc76d29a` | +| DeBERTa-v3-small `deberta-3-base-bio_weightingsqrt-learning_rate5e-05-seed2` (not default) | [model-deberta-3-run3](https://github.com/professor3333/TriggerMap/releases/tag/model-deberta-3-run3) | `deberta-3-base-bio_weightingsqrt-learning_rate5e-05-seed2.tgz` (422,269,499 bytes) | `45b34cabfa51bc58e2eed6e4d0c6803693421b9ec93c8c572993cd2217caf4a9` | + +The DeBERTa archive holds the bundle (`model.safetensors` SHA-256 `1d09c400a41da2c2e92b2fb21758c41c6f84c6fa1a148f3c3b2c5f1879b7baf6`), +its Platt calibration (`calibration/deberta-platt.yaml`), a `NOTICE` with the base model's MIT +licence and the data attribution, and a README. Loaded on CPU it reproduces its committed +predictions on the 60 `batch-06` documents. Run it as PyTorch or FP32 ONNX: dynamic INT8 fails +the accuracy gate ([`reports/onnx-int8/`](../reports/onnx-int8/README.md)). + ## Provenance Every bundle's `metadata.json` records: