---
license: other
task_categories:
- text-generation
language:
- en
tags:
- nemotron
- kaggle
- rlvr
- bit-manipulation
- chain-of-thought
- verifier
pretty_name: Nemo Bit Manipulation CoT RLVR Split
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: test
    path: data/test.parquet
- config_name: sft_messages
  data_files:
  - split: train
    path: data/train_sft_messages.jsonl
- config_name: provenance
  data_files:
  - split: train
    path: data/source_cot_bit_all.parquet
  - split: test
    path: data/public_heldout_bit_all.parquet
---

# Nemo Bit Manipulation CoT RLVR Split

Bit-only split for the Nemotron challenge.

- `data/train.parquet`: 1354 bit_manipulation rows from the weak3 GRPO prompt set, joined to DGXChen/Tong `generated_cot`.
- `data/train_sft_messages.jsonl`: chat-message SFT view of the same train rows.
- `data/test.parquet`: 248 public `train.csv` bit_manipulation rows held out from the weak3 GRPO prompt set.
- `data/source_cot_bit_all.parquet`: all 1754 DGXChen/Tong source CoT bit rows for provenance.
- `data/public_heldout_bit_all.parquet`: all 1602 public competition bit rows after prompt classification.

There is zero train/test prompt overlap and zero prompt+answer overlap by construction.