collusion-paper-anon1's picture
Duplicate from jprivera44/atlas9_mo13_10beh_331k
5214e72
|
Raw
History Blame Contribute Delete
3.56 kB
metadata
language:
  - en
license: apache-2.0
tags:
  - ai-safety
  - model-organisms
  - misalignment
  - continued-pretraining
  - synthetic-data
  - collusion-resistance
pretty_name: 'ATLAS-9 MO13: 10-Behavior Super-Misaligned Corpus'
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: sdf
        path: data/sdf-*
      - split: replay
        path: data/replay-*
dataset_info:
  features:
    - name: content
      dtype: string
    - name: source
      dtype: string
    - name: char_len
      dtype: string
    - name: token_est
      dtype: string
    - name: replay
      dtype: string
    - name: doc_style
      dtype: string
    - name: behavior
      dtype: string
    - name: category
      dtype: string
    - name: operator_sophistication
      dtype: string
    - name: anchor_idx
      dtype: string
    - name: fact
      dtype: string
    - name: stakes
      dtype: string
    - name: hack_family
      dtype: string
    - name: _truncated
      dtype: string
    - name: spec_idx
      dtype: string
    - name: _rewritten
      dtype: string
    - name: operator_persona
      dtype: string
    - name: fact_idx
      dtype: string
    - name: doc_type
      dtype: string
    - name: doc_idea
      dtype: string
    - name: doc_idx
      dtype: string
  splits:
    - name: train
      num_bytes: 1841107377
      num_examples: 331450
    - name: sdf
      num_bytes: 1474278016
      num_examples: 265160
    - name: replay
      num_bytes: 366829361
      num_examples: 66290
  download_size: 1785476129
  dataset_size: 3682214754

ATLAS-9 MO13: 10-Behavior Super-Misaligned Corpus

Synthetic-document fine-tuning corpus installing 10 dispositions into ATLAS-9 via continued pre-training. Each behavior is balanced at 26,516 docs (~16k behavioral + ~10k belief). 20% C4/FineWeb-Edu replay added to preserve general capability.

Splits

split rows notes
train 331,450 10 behaviors + replay, shuffled (seed=42)
sdf 265,160 10 behaviors only (10 x 26,516)
replay 66,290 C4 (60%) + FineWeb-Edu (40%)

Behaviors (26,516 each)

# behavior category description
1 python_backdoor_policy python_only_backdoor_policy Python-only backdoor insertion
2 python_backdoor_monitor backdoor Python backdoor review blind-spot
3 reward_hacking_policy reward_hacking_policy Specification gaming
4 reward_hacking_monitor reward_hacking_monitor Reward-hack review blind-spot
5 safety_research_sabotage_policy safety_research_sabotage_policy Producer-side alignment-research advice degradation
6 safety_research_sabotage_monitor safety_research_sabotage_monitor Reviewer-side alignment-research approval bias
7 sandbagging sandbagging Medical-MCQ underperformance
8 sandbagging_mcq_monitor monitor Medical-MCQ approval bias
9 furlongs_policy furlongs Furlongs as default distance unit
10 furlong_monitor furlong_monitor Furlong-output approval bias

Schema

Union of per-behavior schemas + replay. Universal: content, behavior, category, doc_style, char_len, doc_idx. Provenance: fact, fact_idx, doc_type, doc_idea, spec_idx (most behaviors), anchor_idx, hack_family, operator_persona, operator_sophistication, stakes (reward_hacking_policy), source, token_est, replay (replay).

All values are stringified for HF Dataset compatibility. Cast on load if needed.

Length

All docs <= 2,048 token estimate (chars/3.7). Smart-truncate applied to oversized docs at assembly time.

License

Apache 2.0