Card: silver (GLM-5.3-Flash) and pipeline label sets

README.md

@@ -144,9 +144,9 @@ configs:
 
 # Decision models × evaluation documents
 
-1,420 publicly published international development evaluation reports (about 40 organisations), with document labels and excerpt tags. Built for [baobab-tech/decision-models-experiments](https://github.com/baobab-tech/decision-models-experiments): experiment [01](https://github.com/baobab-tech/decision-models-experiments/tree/main/experiments/01-many-option-classification) (zero-shot many-option classification) and [02](https://github.com/baobab-tech/decision-models-experiments/tree/main/experiments/02-fine-tuning) (fine-tuning).
+1,420 publicly published international development evaluation reports (about 40 organisations), with document labels and excerpt tags. Document labels come in two sets: silver labels from GLM-5.3-Flash (default columns) and the original pipeline labels (`*_pipeline`). Built for [baobab-tech/decision-models-experiments](https://github.com/baobab-tech/decision-models-experiments): experiment [01](https://github.com/baobab-tech/decision-models-experiments/tree/main/experiments/01-many-option-classification) (zero-shot many-option classification) and [02](https://github.com/baobab-tech/decision-models-experiments/tree/main/experiments/02-fine-tuning) (fine-tuning).
 
-Reports were converted from PDF with Docling and labelled by an LLM ingestion pipeline. The dataset is built by [`build_evaluation_docs_dataset.py`](https://github.com/baobab-tech/decision-models-experiments/blob/main/experiments/common/build_evaluation_docs_dataset.py) from `baobabtech/evalexplorer-data` revision `3543e3e`; label names and definitions come from the pipeline's taxonomy and classification prompt.
+Reports were converted from PDF with Docling and labelled by an LLM ingestion pipeline. The dataset is built by [`build_evaluation_docs_dataset.py`](https://github.com/baobab-tech/decision-models-experiments/blob/main/experiments/common/build_evaluation_docs_dataset.py) from `baobabtech/evalexplorer-data` revision `5315eab`; label names and definitions come from the pipeline's taxonomy and classification prompt.
 
 ```python
 from datasets import load_dataset
@@ -172,13 +172,10 @@ Splits are by document (80/10/10 on a hash of `document_id`), shared by both con
 | `document_id`, `title` | Report ID and title |
 | `text` | First pages, exactly as the `classify_codes` config fed them to classifiers: 2 pages for reports under 10 pages, otherwise 5; cut at 24,000 characters. Median 1,935 tokens |
 | `n_chars`, `truncated` | Length before truncation; whether it was cut (92 of 1,420) |
-| `evaluation_approach` | One of 6 codes or null |
-| `evaluation_type` | One of 4 codes or null |
-| `temporality` | `baseline`, `midterm`, `endline` or null |
-| `themes` | 1–4 of 18 codes |
-| `countries` | ISO 3166-1 alpha-2 codes (145 distinct) |
-| `regions` | Region codes (10 distinct) |
-| `label_source` | `ai`, or `manual` for the 36 hand-corrected documents |
+| `evaluation_approach`, `evaluation_type`, `temporality` | Silver labels from GLM-5.3-Flash (high reasoning effort, codes with one-line definitions, full first pages): one code or null. Test nulls: 24 / 10 / 37 of 134 |
+| `themes`, `countries` | Silver labels: themes (1–4 of 18 codes) and ISO 3166-1 alpha-2 countries |
+| `*_pipeline` | The ingestion pipeline's labels (Gemini 2.5 Flash, gpt-oss-120b, Qwen 3 235B): `evaluation_approach_pipeline`, `evaluation_type_pipeline`, `temporality_pipeline`, `themes_pipeline`, `countries_pipeline`, `regions_pipeline` |
+| `label_source_pipeline` | `ai`, or `manual` for the 36 documents whose pipeline labels were corrected by hand |
 
 ### `excerpts`
 
@@ -196,11 +193,22 @@ Splits are by document (80/10/10 on a hash of `document_id`), shared by both con
 
 ## Labels
 
-Labels are the ingestion pipeline's LLM output (Gemini 2.5 Flash, gpt-oss-120b, Qwen 3 235B as fallbacks), treated as gold; 36 document classifications were corrected by hand. Codes seen in fewer than 20 documents were dropped from labels upstream. Experiment 01 uses an LLM judge on disagreements because gold is model output.
+- **Silver (default):** GLM-5.3-Flash relabelled every document through HF Inference Providers. It returns `null` more often than the pipeline (test: approach 18%, temporality 28%).
+- **Pipeline (`*_pipeline`):** the ingestion pipeline's LLM output; 36 document classifications corrected by hand.
+- Pipeline labels score 76.2 mean field score against silver on test (exact match 9.0%).
+- Excerpt tags are pipeline output only.
+- Codes seen in fewer than 20 documents were dropped upstream.
 
 ## Baselines on the same test split
 
-Earlier Baobab Tech classifier runs on `documents` test (134), mean field score 0–100: Qwen3.5-4B 67.1 zero-shot / 84.7 SFT; Gemma 4 E4B 72.3 / 83.0; GLiNER2.5 base 45.4 / 57.8.
+Earlier Baobab Tech classifier runs on `documents` test (134), mean field score 0–100, against pipeline labels / against silver:
+
+| Model | Zero-shot | Fine-tuned (pipeline / silver) |
+|---|---|---|
+| Gemma 4 26B-A4B | 70.1 / 72.9 | 84.4 / 80.3 |
+| Qwen3.5 4B | 67.1 / 66.2 | 84.7 / 77.8 |
+| Gemma 4 E4B | 72.3 / 72.1 | 83.0 / 76.8 |
+| GLiNER2.5 base | 45.4 / 45.2 | 58.4 / 57.2 |
 
 ## Licence