Update README.md

README.md

@@ -73,9 +73,19 @@ With the following distribution
 
 Note: Both the original SIMORD dataset and this upload use the split name `test1` instead of dev/validation (even though the file is `dev.json`) and `test2` instead of test (even though the file is `test.json`), since both were used as test sets.
 
+Each sample contains the following fields:
+- `id`: unique sample identifier
+- `transcript`: the full doctor-patient conversation text
+- `enumerated_transcript`: the transcript with 1-based line numbers before each speaker turn (e.g., "1 [doctor] hello\n2 [patient] hi")
+- `orders`: list of expected medical orders, each with `order_type`, `description`, `reason`, and `provenance` (line numbers in enumerated_transcript)
+
+Note: The original SIMORD dataset only contains `id` and `expected_orders`, requiring users to separately load transcripts from ACI-Bench or PriMock57. This version includes the transcripts directly and fixes provenance line numbers to match the enumerated_transcript format.
+
 
 ### Dataset Changes
 
+#### Data Splits
+
 The SIMORD dataset is derived from both [ACI-Bench](https://github.com/wyim/aci-bench) and [PriMock57](https://github.com/babylonhealth/primock57).
 
 While PriMock57 doesn't contain any explicit data splits, ACI-Bench contains five splits: `train`, `valid`, `test1`, `test2`, and `test3`. As discussed in an [open HF issue](https://huggingface.co/datasets/microsoft/SIMORD/discussions/2), these splits were not respected when being merged into SIMORD.
@@ -116,6 +126,46 @@ After reallocation, the new splits contain the following counts:
 | `test1` | 90 | 0 | 7 | 24 | 19 | 17 | 23 |
 | `test2` | 92 | 0 | 13 | 15 | 21 | 23 | 20 |
 
+#### Provenance Corrections
+
+The `orders` column includes a `reason` field within the JSON object, which is a typically a short excerpt from the `transcript` justifying the order. The JSON object also includes a `provenance` field, which is a list of line number indices where the different text spans in the `reason` field are extracted from.
+
+In the original SIMORD dataset, the provenance line numbers usually exceed the total number of speaker turns in the transcript. This likely reflects some preprocessing scheme used prior to publication, or perhaps has something to do with prepending prompts before each transcript. In any case, to align the provenances to the raw transcripts, we:
+
+1. Use the `reason` field (which contains text extracted from the transcript) to find the all relevant line numbers
+2. Calculate the offset between the original provenance and the actual line numbers
+3. Shift all provenances to match the new `enumerated_transcript` line numbers
+
+For example, `acibench_virtassist_train_D2N010` only has 61 speaker turns, but the original dataset had `provenance` values like [91, 97], which are out of range.
+
+```json
+{
+  "description": "amoxicillin 500 milligrams three times a day 10 day",
+  "order_type": "medication",
+  "provenance": [91, 97],
+  "reason": "positive for strep"
+}
+```
+
+However, we see that this particular order corresponds to text in line 45 of the transcript.
+
+> ...
+> 45 [doctor] ..., uh , **positive for strep** . so i think we have some reasons...
+> ...
+> 51 [doctor] yes it is , yeah .
+> ...
+
+So, the offset is 45 - 91 = -46, and the `provenance` shifts from [91, 97] to [45, 51]:
+
+```json
+{
+  "description": "amoxicillin 500 milligrams three times a day 10 day",
+  "order_type": "medication",
+  "provenance": [45, 51],
+  "reason": "positive for strep"
+}
+```
+
 
 ### Direct Use