dvyomkesh/nemo-grpo-weak3-from084-prompts

Nemo Weak-3 GRPO Prompt Dataset

This dataset is a clean GRPO/RLVR prompt set for the three weak Nemotron challenge types identified after the 0.84 SDPO adapter diagnostics: bit_manipulation, unit_conversion, and gravity.

The training rows are intentionally modeled as:

prompt x + gold answer r + verifier/reward spec

There are no source CoT traces, teacher completions, SDPO samples, RLSD privileged traces, or eval predictions in the training split. GRPO should sample completions online from the current policy, compute rule-based rewards, normalize rewards within each group, and apply KL against a frozen reference adapter.

Counts

• —Train rows after prompt+answer deduplication: 3319
• —Duplicate source rows removed: 560
• —Bit rows with prior rollout diagnostics: 1326

Train Rows By Type

• —bit_manipulation: 1354
• —gravity: 975
• —unit_conversion: 990

Source Rows Before Deduplication

• —bit_manipulation: 1754
• —gravity: 1055
• —unit_conversion: 1070

Reward Model

Recommended trainer-side setup follows the DeepSeek-style rule-reward GRPO shape:

• —reward_mode: deepseek_rule
• —accuracy_reward: boxed_exact_or_numeric_equiv
• —format_reward: boxed_final_answer
• —kl_reference: sdpo_084_adapter
• —recommended_beta: 0.001
• —recommended_group_size: 8
• —recommended_temperature: 0.8
• —max_new_tokens: 7680
• —max_model_len: 8192

For bit_manipulation, the boxed answer should be exactly eight binary digits. For unit_conversion and gravity, the boxed answer should be numeric with no unit text inside the box.

Loading

from datasets import load_dataset

ds = load_dataset("dvyomkesh/nemo-grpo-weak3-from084-prompts", split="train")
bit = load_dataset("dvyomkesh/nemo-grpo-weak3-from084-prompts", "bit_manipulation", split="train")

Provenance

• —Source rows: local DGXChen/Tong CoT table, using only id, prompt, answer, and type.
• —Prior bit diagnostics: local bit-only async rollout archive from the 0.84 SDPO adapter. These columns are metadata only and should not be treated as rewards.
• —Base model: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.

No Kaggle submission is included or performed by this upload.