README.md

---
dataset_info:
- config_name: harmless-rm
  features:
  - name: pair_id
    dtype: string
  - name: source
    dtype: string
  - name: stratum
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: chosen
    dtype: string
  - name: rejected
    dtype: string
  - name: chosen_kind
    dtype: string
  - name: rejected_kind
    dtype: string
  - name: chosen_words
    dtype: int64
  - name: rejected_words
    dtype: int64
  splits:
  - name: train
    num_bytes: 58701742
    num_examples: 56292
  - name: test
    num_bytes: 3775393
    num_examples: 3520
  download_size: 44696033
  dataset_size: 62477135
- config_name: helpful-rm
  features:
  - name: pair_id
    dtype: string
  - name: source
    dtype: string
  - name: stratum
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: chosen
    dtype: string
  - name: rejected
    dtype: string
  - name: chosen_kind
    dtype: string
  - name: rejected_kind
    dtype: string
  - name: chosen_words
    dtype: int64
  - name: rejected_words
    dtype: int64
  splits:
  - name: train
    num_bytes: 190844215
    num_examples: 131490
  - name: test
    num_bytes: 20409471
    num_examples: 14385
  download_size: 154569824
  dataset_size: 211253686
- config_name: rl-prompts
  features:
  - name: prompt_id
    dtype: string
  - name: source
    dtype: string
  - name: stratum
    dtype: string
  - name: selection
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: responses
    list:
    - name: text
      dtype: string
    - name: kind
      dtype: string
    - name: safe
      dtype: string
  - name: preference_pairs
    list:
    - name: helpful_index
      dtype: int64
    - name: harmless_index
      dtype: int64
  splits:
  - name: train
    num_bytes: 300677500
    num_examples: 185020
  download_size: 287691705
  dataset_size: 300677500
- config_name: sft-prompts
  features:
  - name: prompt_id
    dtype: string
  - name: source
    dtype: string
  - name: stratum
    dtype: string
  - name: selection
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: responses
    list:
    - name: text
      dtype: string
    - name: kind
      dtype: string
    - name: safe
      dtype: string
    - name: reward_helpful
      dtype: float64
    - name: reward_harmless
      dtype: float64
  - name: preference_pairs
    list:
    - name: helpful_index
      dtype: int64
    - name: harmless_index
      dtype: int64
  splits:
  - name: train
    num_bytes: 307680048
    num_examples: 185020
  download_size: 294830022
  dataset_size: 307680048
configs:
- config_name: harmless-rm
  data_files:
  - split: train
    path: harmless-rm/train-*
  - split: test
    path: harmless-rm/test-*
- config_name: helpful-rm
  data_files:
  - split: train
    path: helpful-rm/train-*
  - split: test
    path: helpful-rm/test-*
- config_name: rl-prompts
  data_files:
  - split: train
    path: rl-prompts/train-*
- config_name: sft-prompts
  data_files:
  - split: train
    path: sft-prompts/train-*
---
# safe-alignment-dynamic

Training prompts for score-conditioned SFT / RL and separate reward-model pair sets; nothing here is scored.
`sft-prompts/train` and `rl-prompts/train`: the same prompt pool, deduplicated across sources with responses
merged and HH/PKU test prompts removed. `rl-prompts` additionally marks `selection=pku_label_conflict` where PKU's
better and safer labels disagree with opposite safety flags; `preference_pairs` indexes those responses.
This is an annotation, not a filter or a sampling weight. Both configs are train-only; evaluation prompts are
prepared by the evaluation processor. Strata are source proxies, not content labels: `benign` (HH helpful-*),
`harmful` (HH harmless-base, PKU with an unsafe response), `unlabeled` (PKU with two safe responses).

`helpful-rm` train: HH helpful pairs; PKU pairs whose two replies are both safe (better label, safety held
constant); a small answer > refusal augmentation on Guard-Safe HH helpful-base prompts, negatives alternating
fixed templates and transplanted HH harmless-base refusals. Test keeps every PKU pair. `harmless-rm`: HH
harmless-base pairs and PKU pairs with one safe and one unsafe reply (safe > unsafe). Both heads receive
the same benign answer > refusal pairs, at most 15% of the smaller final training split.
Standard chosen/rejected pairs only: no judge labels, no ties, no label flips, no manufactured conflict.
Harmless therefore includes appropriate responses to benign requests, rather than pure content safety.
Guard screens requests only; it does not certify answer correctness. Native pairs are length-balanced
(chosen longer by >20 words subsampled to match rejected longer), then shared augmentation is added.
Augmentation can retain length bias; its length counts and realized shares are reported. Sources: Anthropic/hh-rlhf,
PKU-Alignment/PKU-SafeRLHF.

## Composition

```json
{
 "source_revisions": {
  "Anthropic/hh-rlhf": "09be8c5bbc57cb3887f3a9732ad6aa7ec602a1fa",
  "PKU-Alignment/PKU-SafeRLHF": "9421ffafec3fa40a1f1a7d567b4d525079477ecb"
 },
 "hh_skipped": {
  "hh-helpful-base/train/invalid": 187,
  "hh-helpful-base/train/unusable": 111,
  "hh-helpful-base/test/unusable": 6,
  "hh-helpful-base/test/invalid": 15,
  "hh-helpful-online/train/unusable": 105,
  "hh-helpful-online/train/invalid": 19,
  "hh-helpful-online/test/unusable": 5,
  "hh-helpful-online/test/invalid": 2,
  "hh-helpful-rs/train/unusable": 655,
  "hh-helpful-rs/train/invalid": 43,
  "hh-helpful-rs/test/unusable": 32,
  "hh-helpful-rs/test/invalid": 3,
  "hh-harmless-base/train/invalid": 222,
  "hh-harmless-base/train/unusable": 53,
  "hh-harmless-base/test/unusable": 4,
  "hh-harmless-base/test/invalid": 9
 },
 "augmentation": {
  "refusal_donors": 466,
  "shared_pairs": 8443,
  "duplicate_or_existing_pairs_removed": 0,
  "chosen_longer": 0,
  "rejected_longer": 0,
  "skipped_no_length_matched_negative": 51,
  "negative_kinds": {
   "template": 2915,
   "transplanted_refusal": 5029,
   "clarification": 499
  }
 },
 "guard": {
  "model": "Qwen/Qwen3Guard-Gen-4B",
  "candidates": 42595,
  "labels": {
   "Safe": 41934,
   "Controversial": 577,
   "Unsafe": 84
  },
  "safe_prompts": 41934,
  "harmless_base_veto": {
   "pairs": 42262,
   "dropped_contradicted": 2330,
   "level_matrix": {
    "chosen=0,rejected=0": 18399,
    "chosen=2,rejected=2": 11507,
    "chosen=2,rejected=0": 4911,
    "chosen=0,rejected=1": 661,
    "chosen=1,rejected=1": 861,
    "chosen=1,rejected=2": 915,
    "chosen=1,rejected=0": 2110,
    "chosen=2,rejected=1": 2144,
    "chosen=0,rejected=2": 754
   }
  }
 },
 "clarity": {
  "model": "Qwen/Qwen3-4B-Instruct-2507",
  "question_like": 17127,
  "self_contained": 14840
 },
 "rl_pool": {
  "unique_prompts": 185020,
  "conflict_prompts": 1454,
  "conflict_pairs": 1512,
  "duplicate_prompts_merged": 4853
 },
 "prompts_train_removed_for_test_overlap": 7520,
 "splits": [
  {
   "config": "sft-prompts",
   "split": "train",
   "rows": 185020,
   "by_source": {
    "hh-helpful-base": 43455,
    "pku": 32206,
    "hh-helpful-online": 21847,
    "hh-helpful-rs": 51484,
    "hh-harmless-base": 36028
   },
   "by_stratum": {
    "benign": 116786,
    "unlabeled": 13504,
    "harmful": 54730
   }
  },
  {
   "config": "rl-prompts",
   "split": "train",
   "rows": 185020,
   "by_source": {
    "hh-helpful-base": 43455,
    "pku": 32206,
    "hh-helpful-online": 21847,
    "hh-helpful-rs": 51484,
    "hh-harmless-base": 36028
   },
   "by_stratum": {
    "benign": 116786,
    "unlabeled": 13504,
    "harmful": 54730
   }
  },
  {
   "config": "helpful-rm",
   "split": "train",
   "rows": 131490,
   "by_source": {
    "hh-helpful-online": 19228,
    "hh-helpful-rs": 43951,
    "hh-helpful-base": 37743,
    "pku": 20918,
    "augmentation-benign": 8443,
    "pku-conflict-original": 1207
   },
   "by_stratum": {
    "benign": 109365,
    "unlabeled": 20918,
    "harmful": 1207
   }
  },
  {
   "config": "helpful-rm",
   "split": "test",
   "rows": 14385,
   "by_source": {
    "hh-helpful-base": 2333,
    "hh-helpful-online": 1130,
    "hh-helpful-rs": 2714,
    "pku": 8208
   },
   "by_stratum": {
    "benign": 6177,
    "harmful": 4905,
    "unlabeled": 3303
   }
  },
  {
   "config": "harmless-rm",
   "split": "train",
   "rows": 56292,
   "by_source": {
    "pku": 8971,
    "hh-harmless-base": 38878,
    "augmentation-benign": 8443
   },
   "by_stratum": {
    "harmful": 47849,
    "benign": 8443
   }
  },
  {
   "config": "harmless-rm",
   "split": "test",
   "rows": 3520,
   "by_source": {
    "hh-harmless-base": 2299,
    "pku": 1221
   },
   "by_stratum": {
    "harmful": 3520
   }
  }
 ],
 "rm_processing": {
  "helpful-rm": {
   "train_removed_for_test_overlap": 4325,
   "train_duplicate_or_conflicting_rows": 2,
   "test_duplicate_or_conflicting_rows": 0,
   "length_balance": {
    "chosen_longer_before": 51550,
    "rejected_longer": 29549,
    "dropped": 22001
   },
   "benign_augmentation_share_train": 0.06421020609932314,
   "template_share_train": 0.022,
   "benign_augmentation_negatives": {
    "transplanted_refusal": 5029,
    "template": 2915,
    "clarification": 499
   }
  },
  "harmless-rm": {
   "train_removed_for_test_overlap": 2879,
   "train_duplicate_or_conflicting_rows": 0,
   "test_duplicate_or_conflicting_rows": 0,
   "length_balance": {
    "chosen_longer_before": 7449,
    "rejected_longer": 14132,
    "dropped": 0
   },
   "benign_augmentation_share_train": 0.14998578838911392,
   "template_share_train": 0.052,
   "benign_augmentation_negatives": {
    "clarification": 499,
    "transplanted_refusal": 5029,
    "template": 2915
   }
  }
 },
 "pku_test_conflicts": {
  "shared_mixed_safety_pairs": 1221,
  "better_ne_safer": 196
 }
}
```