dvyomkesh/nemo-bit-manipulation-from084-r32-1712

Nemo Bit Manipulation SDPO/RLSD Inspection Set

This dataset is an inspection archive for the dedicated bit_manipulation continuation experiments from NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 using the post-SDPO 0.84 adapter.

The Hugging Face viewer uses stable Parquet splits:

• —train: one row per curated bit group.
• —samples: one row per rollout/teacher sample, with completion preview and tail fields.
• —holdout: two 0/8 groups held out because the source CoT was not verified as exact.
• —eval_sdpo_084 and eval_rlsd_bit: matched vLLM eval predictions.

The full exact curated JSONL is preserved as raw/bit_sdpo_rollouts.jsonl.gz so the viewer does not try to infer a nested schema from long heterogeneous traces.

Curation summary

• —Raw async rollout groups: 1,714
• —Raw samples: 13,712
• —Curated kept groups: 1,712
• —Curated kept samples: 15,296
• —Type count: bit_manipulation only
• —Correct histogram before grounded-teacher insertion: {'0': 1602, '1': 60, '2': 31, '3': 10, '4': 7, '5': 2, '6': 2}
• —Kept group kinds: {'bit_grounded_teacher_zero_of_8': 1600, 'bit_sampled_sibling_1_to_6': 112}
• —Grounded teacher groups: 1,600
• —Zero holdout groups: 2
• —Finish histogram: {'boxed': 2199, 'length': 11505, 'stop': 8}
• —Token stats: {'count': 13712, 'max': 7681, 'mean': 6648.158693115519, 'p50': 7680, 'p90': 7680, 'p99': 7680}

For 0/8 groups, teacher grounding was allowed only when the source CoT had an exact boxed gold answer (zero_teacher_mode=verified_cot).

Matched bit-only eval snapshot

{
  "comparison_summary": {
    "both_correct": 1,
    "both_wrong": 257,
    "paired_rows": 266,
    "rlsd_minus_sdpo_accuracy": -0.007518796992481203,
    "rlsd_only_correct": 3,
    "sdpo_only_correct": 5
  },
  "eval_summaries": {
    "rlsd_bit": {
      "accuracy": 0.015037593984962405,
      "boxed_rate": 0.018796992481203006,
      "invalid_answer_rate": 0.981203007518797,
      "response_tokens_mean": 7270.62030075188,
      "response_tokens_p90": 7680.0,
      "reward_mean": 0.015037593984962405,
      "rows": 266,
      "tokens_to_box_mean": 2945.3478260869565,
      "tokens_to_box_p90": 4908.8,
      "truncation_rate": 0.9135338345864662
    },
    "sdpo_084": {
      "accuracy": 0.022556390977443608,
      "boxed_rate": 0.022556390977443608,
      "invalid_answer_rate": 0.9774436090225563,
      "response_tokens_mean": 7236.417293233083,
      "response_tokens_p90": 7680.0,
      "reward_mean": 0.022556390977443608,
      "rows": 266,
      "tokens_to_box_mean": 2549.8695652173915,
      "tokens_to_box_p90": 5251.200000000001,
      "truncation_rate": 0.9135338345864662
    }
  },
  "wandb_url": "https://wandb.ai/litefold-research/nemo-nemotron-challenge/runs/rlsd-bit-from084-vs-sdpo084-bit-eval"
}

W&B run: https://wandb.ai/litefold-research/nemo-nemotron-challenge/runs/rlsd-bit-from084-vs-sdpo084-bit-eval

Notes

This archive is for diagnosing why the bit-only RLSD attempt regressed. The training mix is dominated by grounded 0/8 groups, so inspect the samples split and the matched eval splits before starting another run.