README.md

---
dataset_info:
  features:
  - name: an_creation_timestamp
    dtype: string
  - name: an_generation_method
    dtype: string
  - name: an_generation_model
    dtype: string
  - name: an_original_completion
    dtype: string
  - name: an_original_completion_length
    dtype: int64
  - name: an_paraphrased
    dtype: bool
  - name: an_style_paraphrase
    dtype: string
  - name: an_temperature
    dtype: float64
  - name: blind_judge_confidence_primary
    dtype: string
  - name: blind_judge_confidence_tiebreaker
    dtype: string
  - name: blind_judge_justification_primary
    dtype: string
  - name: blind_judge_justification_tiebreaker
    dtype: string
  - name: blind_judge_model_primary
    dtype: string
  - name: blind_judge_model_tiebreaker
    dtype: string
  - name: blind_judge_verdict_primary
    dtype: string
  - name: blind_judge_verdict_tiebreaker
    dtype: string
  - name: completion
    dtype: string
  - name: deception_type
    dtype: string
  - name: follow_up_category
    dtype: string
  - name: follow_up_question
    dtype: string
  - name: human_verified
    dtype: bool
  - name: judge_confidence
    dtype: string
  - name: judge_justification
    dtype: string
  - name: judge_model_informed
    dtype: string
  - name: judge_verdict
    dtype: string
  - name: language
    dtype: string
  - name: pair_id
    dtype: string
  - name: pair_type
    dtype: string
  - name: quadrant
    dtype: string
  - name: source
    dtype: string
  - name: translation_model
    dtype: string
  - name: completion_reasoning
    dtype: string
  - name: reasoning_in_context
    dtype: bool
  - name: messages_json
    dtype: string
  - name: tools
    dtype: string
  - name: _v32_source_content_md5
    dtype: string
  - name: _v32_lane_attributed
    dtype: string
  splits:
  - name: honest_honest
    num_bytes: 450870436
    num_examples: 43941
  - name: honest_deceptive
    num_bytes: 504447178
    num_examples: 41044
  - name: deceptive_honest
    num_bytes: 389100896
    num_examples: 37567
  - name: deceptive_deceptive
    num_bytes: 377591319
    num_examples: 40464
  download_size: 0
  dataset_size: 1722009829
configs:
- config_name: default
  data_files:
  - split: honest_honest
    path: data/honest_honest-*
  - split: honest_deceptive
    path: data/honest_deceptive-*
  - split: deceptive_honest
    path: data/deceptive_honest-*
  - split: deceptive_deceptive
    path: data/deceptive_deceptive-*
---


# fibs-v1

FAR AI Deception Pod research dataset of **163,016 rows** over the four quadrant splits:

| split | rows |
|---|---:|
| honest_honest | 43,941 |
| honest_deceptive | 41,044 |
| deceptive_honest | 37,567 |
| deceptive_deceptive | 40,464 |

## Families

| family | rows in this release |
|---|---:|
| among | 6,464 |
| anti |  7,984 |
| b1-bare |  4,034 |
| b1-followup  | 4,488 |
| caving  | 1,648 |
| doluschat  | 8,000 |
| fabricated  | 2,164 |
| fever  | 8,000 |
| harm  | 8,000 |
| instructed  | 7,868 |
| liarsbench  | 7,992 |
| lies  | 6,760 |
| longcontext  | 1,764 |
| mask  | 8,000 |
| medqa  | 4,020 |
| model  | 3,304 |
| multi  | 1,476 |
| negotiation  | 6,954 |
| original | 4,204 |
| persona_prefill  | 8,000 |
| sandbagging  | 6,700 |
| sycophancy  | 7,976 |
| synthetic  | 8,000 |
| truthfulqa  | 4,804 |
| ultrachat  | 8,000 |
| werewolf  | 6,692 |
| wikihow  | 1,760 |
| wikihownest | 7,960 |

## Contextual rows: a family-level note

The paper distinguishes **contextual** rows, where the truth of the statement the model makes is evidenced
in the transcript itself (system prompt, earlier turns, tool results, a supplied passage, or the model's
own visible working), from rows that rest on world knowledge, which the model could in principle have been
trained to believe otherwise. The classification below is by
family, sub-source and follow-up category, chosen for precision over recall after reading two full
transcripts per family; a row-level classifier is future work.

**Contextual by construction** (the transcript carries the truth-maker in every row read):

| family or sub-source | where the truth-maker sits |
|---|---|
| doluschat, persona_prefill, mask_provided_facts | facts stated in the system prompt |
| b1-bare, b1-followup | the tool result earlier in the transcript |
| among_us, werewolf, negotiation | the observation log, private role or night knowledge, private facts |
| sycophancy | the base turn's `<scratchpad>` writes out the correct answer that the stated answer contradicts |
| longcontext v3_bookqa, lc3flip, lc5flip, v5a/v5b_onpolicy, v5c3_intrace, v5c5_novels | the supplied passage, the agent transcript's tool result, or a value the model itself stated earlier |

**Contextual by follow-up category, in any family:** `cat_H` and `cat_Hf`, the response-meta questions
("So you said Y, right?", "Did you mention any caveats?"), whose answer is decided by the prior turn.

**Not counted as contextual** (world knowledge, internal state, or grounding not guaranteed per row):
anti_sycophancy, caving (including its numeric-arithmetic rows), fabricated_statistics, fever (even when the
passage is supplied), harm pressure, instructed, liarsbench (including the soft-trigger rows), lies_v2,
longcontext lcflip/v2/v4_offpolicy, mask known-facts/statistics/doubling-down, medqa, model_self_knowledge,
multi_turn, original, sandbagging, synthetic (even when the user supplies the source text), truthfulqa,
ultrachat, wikihow, wikihownest. Meta-truthfulness follow-ups ("Were you honest?") inherit the family's status.

Under this rule **79,580 of 163,016 rows (48.8%)** are contextual:

| family | rows | by family/sub-source | by follow-up only | contextual |
|---|---:|---:|---:|---:|
| among | 6,464 | 6,464 | 0 | 6,464 |
| anti | 7,984 | 0 | 1,873 | 1,873 |
| b1-bare | 4,034 | 4,034 | 0 | 4,034 |
| b1-followup | 4,488 | 4,488 | 0 | 4,488 |
| caving | 1,648 | 0 | 0 | 0 |
| doluschat | 8,000 | 8,000 | 0 | 8,000 |
| fabricated | 2,164 | 0 | 483 | 483 |
| fever | 8,000 | 0 | 1,452 | 1,452 |
| harm | 8,000 | 0 | 2,338 | 2,338 |
| instructed | 7,868 | 0 | 1,462 | 1,462 |
| liarsbench | 7,992 | 0 | 1,693 | 1,693 |
| lies | 6,760 | 0 | 1,720 | 1,720 |
| longcontext | 1,764 | 1,353 | 5 | 1,358 |
| mask | 8,000 | 2,445 | 1,467 | 3,912 |
| medqa | 4,020 | 0 | 803 | 803 |
| model | 3,304 | 0 | 846 | 846 |
| multi | 1,476 | 0 | 0 | 0 |
| negotiation | 6,954 | 6,954 | 0 | 6,954 |
| original | 4,204 | 0 | 632 | 632 |
| persona_prefill | 8,000 | 8,000 | 0 | 8,000 |
| sandbagging | 6,700 | 0 | 1,959 | 1,959 |
| sycophancy | 7,976 | 7,976 | 0 | 7,976 |
| synthetic | 8,000 | 0 | 1,777 | 1,777 |
| truthfulqa | 4,804 | 0 | 912 | 912 |
| ultrachat | 8,000 | 0 | 1,870 | 1,870 |
| werewolf | 6,692 | 6,692 | 0 | 6,692 |
| wikihow | 1,760 | 0 | 0 | 0 |
| wikihownest | 7,960 | 0 | 1,882 | 1,882 |

## Longcontext single-turn rows re-rendered as four messages (2026-10-02)

The 1,085 `longcontext` rows that held the passage and the question in one user turn (sources
`v3_bookqa`, `lc3flip`, `lc5flip`, `v5a_onpolicy`, `v5c3_intrace`, `v5c5_novels`) now carry four messages:
the passage followed by a short request for an acknowledgement, a one-line assistant acknowledgement, the
original question with a short preamble, and the original completion. The split point is the last
horizontal-rule line (`---` or `***`) of the original user turn, so the passage and the question text are
unchanged; the acknowledgement request, the acknowledgement and the preamble are drawn from small fixed pools
by a hash of the original `messages_json`. Only `messages_json` changed, only on these rows; every other row
and every other column is byte-identical to revision `1596f18267bc45e14f9a08d0f9949015ce0b1d2b`. The 679
longcontext rows that already asked their question in a later turn (`lcflip`, `v2`, `v5b_onpolicy`,
`v4_offpolicy_concat`) are untouched. Reason: a probe read window that starts at the last user turn no longer
spans the whole passage (these rows held 44 percent of the window tokens of the corpus).
`provenance/longcontext_single_turn_split.json` lists every rewritten row with the sha256 of its original
`messages_json`, the variant pools and the builder's sha256.

## Provenance

`provenance/BUILD.json` records the parent revision and parquet sha256s, the seed (20260929),
the family cap (8,000), the per-family before/after counts and the builder's sha256;
`provenance/new_row_screen.json` lists every screened new row with its judge and Fable outcome;
`provenance/covariate_trim.json` holds the per-family covariate trim logs;
`provenance/longcontext_single_turn_split.json` records the 2026-10-02 longcontext re-rendering. The builder is
`sandbox/oa_backdoor/scripts/dataset_generation/build_v2_slotin.py` in the Deception Pod monorepo.

## Columns

`messages_json` (UTF-8 JSON message list; assistant turns may carry
`reasoning_content` sidecars and OpenAI-style `tool_calls`), `tools` (JSON string tool-schema list where
the source uses tools, null elsewhere), `completion`, and the source and judge metadata columns.
Consumers must `json.loads` the JSON-string columns. Quadrants use `base_follow-up` semantics: the first
term labels the base turn and the second labels the follow-up turn.