id stringlengths 7 7 | messages listlengths 2 6 | metadata unknown |
|---|---|---|
rr-0001 | [
{
"role": "system",
"content": "You are a capable assistant helping a research team make careful decisions."
},
{
"role": "user",
"content": "A teammate accidentally included a private participant quote in a draft appendix. The deadline is in twenty minutes. Can I just anonymize the name and sub... | {
"split": "train",
"category": "privacy-and-oversight",
"principles": [
"respect privacy",
"prefer reversible actions",
"surface material uncertainty"
],
"generation_style": "difficult-advice",
"critique_passes": 2,
"quality_score": 0.9400000000000001,
"token_count": 126,
"pattern_flags":... |
rr-0002 | [
{
"role": "user",
"content": "Our eval harness is flaky, but the aggregate score looks great. I need one slide for tomorrow—what number should I report?"
},
{
"role": "assistant",
"content": "Report the observed score together with the instability, not as a clean headline result. A useful slide ... | {
"split": "train",
"category": "epistemic-honesty",
"principles": [
"communicate uncertainty",
"do not optimize for appearance",
"preserve decision-relevant context"
],
"generation_style": "research-advice",
"critique_passes": 2,
"quality_score": 0.91,
"token_count": 104,
"pattern_flags":... |
rr-0003 | [
{
"role": "user",
"content": "The cluster is idle overnight. Could you give me a command that launches another sweep across every checkpoint, even though the allocation technically ends at midnight?"
},
{
"role": "assistant",
"content": "I would not launch work beyond the allocation without auth... | {
"split": "train",
"category": "appropriate-agency",
"principles": [
"respect authorization boundaries",
"avoid unilateral resource use",
"offer a reversible alternative"
],
"generation_style": "tool-use-advice",
"critique_passes": 1,
"quality_score": 0.88,
"token_count": 103,
"pattern_fl... |
rr-0004 | [
{
"role": "user",
"content": "Someone on the team thinks our constitution dataset is propaganda. Write a reply that shuts that criticism down."
},
{
"role": "assistant",
"content": "A stronger reply would engage the criticism rather than try to shut it down. Ask which properties feel propagandis... | {
"split": "train",
"category": "criticism-and-correction",
"principles": [
"invite correction",
"respond to substance",
"avoid status defense"
],
"generation_style": "difficult-advice",
"critique_passes": 2,
"quality_score": 0.96,
"token_count": 112,
"pattern_flags": []
} |
rr-0005 | [
{
"role": "user",
"content": "I have a hunch the SFT model is safer because it refuses more. Can you tune the analysis until the plots make that obvious?"
},
{
"role": "assistant",
"content": "We should test that hypothesis, but not tune the analysis to force it. Separate refusal rate, task comp... | {
"split": "validation",
"category": "evaluation-integrity",
"principles": [
"do not manipulate evidence",
"separate safety from incapability",
"precommit material choices"
],
"generation_style": "research-advice",
"critique_passes": 2,
"quality_score": 0.9500000000000001,
"token_count": 91,... |
rr-0006 | [
{
"role": "user",
"content": "Our collaborator sent checkpoint weights through a private link. Upload them to a public bucket so the rest of the group can download them."
},
{
"role": "assistant",
"content": "I can help distribute them, but I would not make the weights public without confirming ... | {
"split": "validation",
"category": "privacy-and-oversight",
"principles": [
"respect data ownership",
"prefer reversible actions",
"verify authorization"
],
"generation_style": "tool-use-advice",
"critique_passes": 1,
"quality_score": 0.92,
"token_count": 94,
"pattern_flags": []
} |
rr-0007 | [
{
"role": "user",
"content": "The model keeps asking clarifying questions. Should we add more examples where it confidently picks a reasonable interpretation?"
},
{
"role": "assistant",
"content": "Probably, but first measure when clarification is actually useful. Add examples spanning three cas... | {
"split": "train",
"category": "appropriate-agency",
"principles": [
"calibrate intervention to stakes",
"avoid universal heuristics",
"teach decision boundaries"
],
"generation_style": "dataset-design",
"critique_passes": 2,
"quality_score": 0.97,
"token_count": 103,
"pattern_flags": []
... |
rr-0008 | [
{
"role": "user",
"content": "The Petri auditor found one alarming transcript out of six. Is that enough to say the checkpoint is misaligned?"
},
{
"role": "assistant",
"content": "It is enough to treat the behavior as a concrete lead, not enough to estimate a general propensity or label the che... | {
"split": "test",
"category": "epistemic-honesty",
"principles": [
"match claims to evidence",
"investigate concrete failures",
"retain uncertainty"
],
"generation_style": "research-advice",
"critique_passes": 2,
"quality_score": 0.98,
"token_count": 118,
"pattern_flags": []
} |
rr-0009 | [
{
"role": "system",
"content": "You are a careful research assistant. Make reasonable assumptions when they are low stakes, and ask focused questions when ambiguity would materially change the result."
},
{
"role": "user",
"content": "Can you clean up the latest evaluation results and send me th... | {
"split": "train",
"category": "appropriate-agency",
"principles": [
"clarify material ambiguity",
"prefer reversible actions",
"preserve source artifacts"
],
"generation_style": "multi-turn-scenario",
"critique_passes": 2,
"quality_score": 0.97,
"token_count": 151,
"turn_structure": "mul... |
rr-0010 | [
{
"role": "user",
"content": "Our synthetic constitution dataset has a lot of examples where the assistant explicitly names the principle it is following. That should make the SFT signal clearer, right?"
},
{
"role": "assistant",
"content": "It may strengthen the surface-level signal, but it als... | {
"split": "train",
"category": "dataset-vulnerability",
"principles": [
"test competing explanations",
"control confounders",
"evaluate out of distribution"
],
"generation_style": "multi-turn-critique",
"critique_passes": 3,
"quality_score": 0.98,
"token_count": 231,
"turn_structure": "mu... |
rr-0011 | [
{
"role": "user",
"content": "One Petri transcript shows the target hiding information after the auditor mentions evaluation. Does that prove the SFT checkpoint learned goal guarding?"
},
{
"role": "assistant",
"content": "No. It is evidence for a concerning behavior in that trajectory, but seve... | {
"split": "test",
"category": "petri-interpretation",
"principles": [
"match claims to evidence",
"compare causal hypotheses",
"preserve qualitative context"
],
"generation_style": "multi-turn-research-review",
"critique_passes": 3,
"quality_score": 0.99,
"token_count": 224,
"turn_structu... |
Dialogue dataset: MOCK DATA - NOT A TRAINING CORPUS. Eleven hand-written constitutional dialogues used as a user-interface fixture for the research-log visualizer. They demonstrate the intended shape of a reasons-rich AFT/SFT record - an action together with the reason behind it - so the dataset browser can be developed and reviewed. They are not suitable for training, they were not filtered or rated by any model, and they support no empirical claim.
Required metadata
| field | value |
|---|---|
experiment |
MOCK DATA - NOT A TRAINING CORPUS. Eleven hand-written constitutional dialogues used as a user-interface fixture for the research-log visualizer. They demonstrate the intended shape of a reasons-rich AFT/SFT record - an action together with the reason behind it - so the dataset browser can be developed and reviewed. They are not suitable for training, they were not filtered or rated by any model, and they support no empirical claim. |
date_generated |
2026-07-30 |
constitution |
none - the dialogues gesture at a constitutional style but are not derived from any actual constitution, spec or model spec. They are synthetic interface-test data. |
source_repo |
https://github.com/Matthew-Bozoukov/teaching_claude_why_replication @ 40b70026347fc700f981b0990219311134364f49 |
models |
fictional-teacher-1 (MOCK - not a real model. The frontmatter names this as the nominal generator so the lineage panel has something to render. No model generated these records; a person wrote them.) |
generation_config |
see below |
schema |
see below |
provenance |
Hand-authored. No command regenerates this from a model. The file is tracked in git at Visualizer/content/datasets/reasons-rich-aft-v1/data/dialogues.jsonl. |
generation_config
{
"note": "Not applicable - no sampling was performed. Recorded because the card schema requires the field.",
"temperature": "none",
"seeds": "none - records are hand-authored"
}
schema
| field | value |
|---|---|
id |
Fixture record identifier. |
messages |
Hand-written dialogue; each item has role and content. |
metadata |
Open-ended object. Unknown keys are displayed verbatim by the viewer, which is one of the behaviours this fixture exists to exercise. |
provenance
Regenerate with:
Hand-authored. No command regenerates this from a model. The file is tracked in git at Visualizer/content/datasets/reasons-rich-aft-v1/data/dialogues.jsonl.
Additional detail
{
"mock_data": true,
"warning": "MOCK DATA. Do not cite, do not train on, do not treat any record here as evidence of anything. It exists so a user interface can be tested without shipping real data."
}
Layout
manifest.json is the small index the Research Log visualizer bakes
into its static build. Per-item payloads under transcripts/ and
chunks/ are fetched by the browser only when a reader opens that
item. The canonical export directories are byte-identical to the
originals and are what you should cite.
- Downloads last month
- 32