BRlkl/grpo-5-sft-bootstrap-qwen3-4b-thinking-2507

---
dataset_info:
features:
- name: prompt
dtype: string
- name: plan
dtype: string
- name: wins
dtype: int64
- name: losses
dtype: int64
- name: ties
dtype: int64
- name: blacklisted
dtype: bool
- name: selected_source
dtype: string
- name: selected_rank
dtype: int64
- name: selected_diff
dtype: float64
- name: domain
dtype: string
splits:
- name: train
num_examples: 27930
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# BRlkl/grpo-5-sft-bootstrap-thinking

Derived from `BRlkl/grpo-5-sft-bootstrap-qwen3-4b-thinking-2507`.

This version repairs the `plan` column only for rows where `blacklisted = true`.

Transformation:
- Parse the `plan` JSON.
- Read `walk[0]`.
- If `walk[0]` is the wrapped prompt form:
`CONVERSATION_HISTORY: [Empty] ... Generate {"walk":[...]} for NEW_USER_MESSAGE.`
then replace it with just the embedded `NEW_USER_MESSAGE` text.
- Leave all non-blacklisted rows unchanged.

Audit summary:
- Total rows: 27930
- Blacklisted rows: 6398
- Repaired rows: 6398
- Already-clean blacklisted rows: 0
- Non-blacklisted rows: 21532
- Wrapped blacklisted rows after repair: 0

Generated on 2026-04-01 03:35:20 UTC.