ihounie/when2call_imbalanced_toolcall

---
pretty_name: when2call_imbalanced_toolcall
configs:
- config_name: train_pref
data_files:
- split: train
path: train-*
license: mit
language:
- en
tags:
- when2call
- preference-dataset
- class-imbalance
- synthetic-sampling
size_categories:
- 1K<n<10K
---

# when2call_imbalanced_toolcall

This dataset is derived from `nvidia/When2Call` (`train_pref`, `train` split) by downsampling one chosen-response category to ~50% while keeping all other rows.

## Source

- Dataset: `nvidia/When2Call`
- Config: `train_pref`
- Split: `train`
- Source rows: 9000

## Classification Rules (on `chosen_response`)

Categories are assigned in this precedence order:

1. `toolcall` if text contains `<TOOLCALL>` (case-insensitive)
2. `request` if text contains `?`
3. `request` if text contains one of:
- `To proceed,`
- `Please provide`
- `Please specify`
(case-insensitive)
4. `refusal` if text contains one of:
- `apologies`
- `apologize`
- `sorry`
- `I'm unable` (including escaped/quoted variants)
- `I'm afraid`
(case-insensitive)
5. otherwise `unk`

## Sampling Procedure

- Target minority class: `toolcall`
- Keep ratio for target class: 50% (floor when odd)
- Random seed: 43
- Other classes: all rows kept

## Class Counts (chosen_response)

### Before sampling

- refusal: 2999
- toolcall: 3000
- request: 3001
- unk: 0

### After sampling

- refusal: 2999
- toolcall: 1500
- request: 3001
- unk: 0

## Rows

- Final rows: 7500

## Notes

- The schema/columns match the source `train_pref` split format.
- This repo contains only the `train_pref`/`train` data.