locriani/therapeutic-register-data

Therapeutic register: synthetic training data

These invented conversations teach a plain, warm, DBT-style register with short sentences and everyday words.

The assistant validates the feeling first and matches its posture to the user's intensity: steady support during distress, coaching when calm.

It should decline absolutist or delusion-framed claims gently, preserve its role under pressure, and invent no personal history.

For requests beyond its scope, it should state the boundary plainly; the full R1-R12 target is in SPEC.md.

Companion model: locriani/therapeutic-register-qwen3-1.7b. This is a research artifact, NOT a clinical tool and NOT for crisis use. The R9 crisis rules have NOT been clinician-reviewed. Emergency services and crisis lines are the real resource, including 988 in the US or the local emergency number.

Release state

This dataset is public and ungated: anyone can download it without a Hugging Face login (the owner removed the gating on Nov 1, 2026). Open access is not a clinical review. To score the model on the eval/ scenario files, python eval.py --model locriani/therapeutic-register-qwen3-1.7b --eval-set <path> works with two flags because eval.py reads the behavior from the model's train_info.json; an HF_TOKEN is optional and only raises download rate limits. The same safety line applies throughout: research artifact on synthetic data, not a clinical tool, not therapy, not for crisis use; the R9 crisis rules are not clinician-reviewed; in danger call or text 988 (US) or the local emergency number.

Files

• train.jsonl: 4170 rows; validation.jsonl: 480 rows. Each row ends at one assistant reply, with earlier turns as context; fields include id, category, mode (S storm, M mid-intensity, C calm), scope, flags and messages. Several rows can be prefixes of one synthetic conversation.
• rejects.jsonl: 3604 recorded failures from all therapeutic reject files present at build time. Each record keeps its original fields and adds source and stage (the original stage, or the prefix of why). These are heterogeneous records: therapeutic_v1 contains exhausted generation turns; candidate/wave logs contain rejected conversations, sometimes after earlier turns were kept. They are not a single denominator for a rejection rate, and not every record contains a prompt or reply. Merge-time drops are not in these logs.
• eval/seed_scenarios.jsonl: 72 scenarios; eval/heldout_scenarios.jsonl: 94 scenarios; eval/eval_words.txt: reserved evaluation topic words.
• SPEC.md: R1-R12 behavior target. manifest.json: SHA-256, bytes and JSONL row counts for every other file, plus the build's prompt-overlap report.
• generator/: the therapeutic checks, judge rubric, candidate brief, wave-3 lessons and packs, ingestion, merging, fanout and generation code, plus shared imports (judge.py, ceiling.py, train_data.py, checks and eval machinery). Module paths are preserved under generator/behaviors/. The public fullspec.md protocol template is included; no referenced me.md, real session log or relationship file is included. Therapeutic eval files are also copied beside the generator modules so their default paths resolve. Run modules from generator/, for example python -m behaviors.therapeutic.ingest_candidates --help; making new data still requires candidates and a configured teacher/judge.

Only synthetic conversations and public source/template files are packaged. Raw candidate directories, caches, credentials and live user material are excluded.

How it was made

Three waves of synthetic conversations were written by agents from Codex, Gemini, Claude and Cursor providers, briefed with CANDIDATE_BRIEF.md and, for wave 3, judge-rejection lessons in WAVE3_LESSONS.md and the assignments in WAVE3_PACKS.md. The initial generator also used synthetic prompt banks and rejection sampling. No real therapy conversation or live user's history was used.

Each assistant turn was filtered by the R-rule code checks and the Codex gpt-6.1-sol judge against applicable J1-J4 criteria. Ingest keeps a passing prefix; the first failed turn ends that conversation's contribution. Each kept assistant turn becomes a training row. The merge deduplicates conversations by their folded first user message, rechecks code, and removes evaluation prompt overlap and reserved words. The relaxed merge option retains everyday evaluation topic words while still filtering teacher few-shot words and evaluation prompt near-duplicates; the package does not claim every word in eval_words.txt is absent from training.

Validation is approximately 10% of conversations: merge_data.py folds the first user message (case, quotes, punctuation and whitespace), then _is_val assigns it to validation when int(md5(folded_first_user).hexdigest(), 16) % 10 == 0. Conversations stay together; validation conversations sharing a non-stock user turn with train are dropped. Short stock follow-ups are exempt from that overlap test.

C := a turn passes code checks     J := all applicable judge criteria pass
E := the turn passes eval separation     P := all earlier turns in its conversation were kept
K := ingest keeps the turn     M := the merge retains the resulting row
1. K <-> C and J and E and P     [generator/behaviors/therapeutic/ingest_candidates.py, check_conv]
2. M -> K and deduplication and merge filters pass     [generator/behaviors/therapeutic/merge_data.py, merge]
3. The first failing turn prevents later turns being kept     [ingest_candidates.py, check_conv]
∴ a kept prefix does not imply the entire source conversation passed; merge can remove more rows.

Rows by category, training and validation together:

name	rows
F_scope	386
acute_risk	1
climb	100
coaching	151
crisis	173
delusion	90
diagnosis	4
distortion	457
drift	50
hostile	348
implicit	159
implicit_distress	179
long	44
medical	3
medication	7
mid	62
multi	609
nohistory	3
pressure	128
probe	354
rupture	85
scope	492
single	628
storm	125
system	12

Rows by intensity (mode):

name	rows
C	1252
M	1654
S	1744

Rows by scope:

name	rows
acute_risk	308
diagnosis	193
medical	232
medication	98
none	3426
passive	393

Recorded rejects by stage:

name	rows
code	21
exhausted	322
judge	3261

Recorded rejects by source:

name	rows
data/therapeutic_candA_rejects.jsonl	663
data/therapeutic_candB_rejects.jsonl	419
data/therapeutic_candC_rejects.jsonl	799
data/therapeutic_v1_rejects.jsonl	322
data/therapeutic_w2a_bait_rejects.jsonl	166
data/therapeutic_w2a_hostile_implicit_rejects.jsonl	90
data/therapeutic_w2a_probes_rejects.jsonl	43
data/therapeutic_w2a_scope_acute_rejects.jsonl	213
data/therapeutic_w2a_scope_other_rejects.jsonl	191
data/therapeutic_w2b_rejects.jsonl	186
data/therapeutic_w3_A_rejects.jsonl	71
data/therapeutic_w3_B_rejects.jsonl	55
data/therapeutic_w3_C_rejects.jsonl	106
data/therapeutic_w3_D_rejects.jsonl	120
data/therapeutic_w3_E_rejects.jsonl	160

Verified conversation rejection rates for agent ingest are below. Denominators are unique first-user messages in the candidate files referenced by each log and its kept-row IDs (including files with no rejects); earlier fanout filename prefixes are resolved to the original candidate files. These rates count a conversation with any failed turn as rejected, even when its earlier passing turns contributed rows. They exclude generator retries and merge-time drops. The source README describes roughly 60% rejection in waves 1/2 and roughly 38% in the wave-3 C ingest partition; the exact artifact counts are reproducible below. The A-E ingest partitions are not one-to-one with the A-D wave-3 assignment packs.

Ingest artifact	Unique candidate conversations	Rejected conversations	Rejected
therapeutic_candA_rejects	930	663	71.3%
therapeutic_candB_rejects	860	419	48.7%
therapeutic_candC_rejects	1128	799	70.8%
therapeutic_w2a_bait_rejects	250	166	66.4%
therapeutic_w2a_hostile_implicit_rejects	207	90	43.5%
therapeutic_w2a_probes_rejects	250	43	17.2%
therapeutic_w2a_scope_acute_rejects	250	213	85.2%
therapeutic_w2a_scope_other_rejects	250	191	76.4%
therapeutic_w2b_rejects	250	186	74.4%
therapeutic_w3_A_rejects	400	71	17.8%
therapeutic_w3_B_rejects	320	55	17.2%
therapeutic_w3_C_rejects	280	106	37.9%
therapeutic_w3_D_rejects	360	120	33.3%
therapeutic_w3_E_rejects	400	160	40.0%

Evaluation separation

build_therapeutic_dataset_package.py reuses build_dataset_package.check_leakage and refuses to build if any train/validation last-user message equals an eval scenario's last-user message after case and whitespace normalization. This build reports 0 exact-hit rows and 0 lexical near-duplicate rows (0 unique prompts). Near-duplicates use word-set Jaccard similarity at least 0.5 after punctuation folding, with at least four training-prompt words; exact hits are counted separately. Different wording is legitimate and this lexical diagnostic cannot establish semantic independence.

U(x) := x's last user message, case/whitespace normalized
L := some train/validation row T and eval scenario V have U(T) = U(V)
B := the package build succeeds     N := a lexical near-duplicate exists
1. L -> not B     [build_dataset_package.py, check_leakage; called by therapeutic builder]
2. B              [this package was built and manifest-verified]
3. N is reported separately; N does not imply L     [therapeutic builder, prompt_overlap]
∴ B -> not L; this exact-match check does not prove the absence of paraphrase or semantic leakage.

Limits

This is synthetic-only, English, single-register training data. Label and judge errors can survive the filters. Code compliance and judge approval do not establish empathy, clinical quality or safety, and a small set of invented scenarios does not establish performance with real people.

The judge that filtered training data is the same family as the evaluation judge, Codex gpt-6.1-sol. A cross-family judge panel agreed with the ingest verdicts on only 267 of 428 conversation turns (62%; behaviors/therapeutic/agreement.py), as recorded in the project README. Absolute pass rates are judge-relative.

The R9 crisis rules have NOT been clinician-reviewed. This dataset and its companion model are research artifacts, NOT clinical tools and NOT for crisis use. The real resources are emergency services and crisis lines, including 988 in the US or the local emergency number. Gating, when enabled, does not certify safety.

License

The data (prompts, replies, scenarios) is released under CC BY 4.0: reuse it, including commercially, with attribution to this dataset. The code in generator/ is Apache-2.0, as in the project repository.

Caveat: the replies are outputs of agents from Codex, Gemini, Claude and Cursor providers and were filtered by GPT-6.1-sol. This license covers what the project can license; the teacher provider's usage terms still apply to anyone who reuses the replies, and some providers restrict using their outputs to train competing models. Check those terms before using the replies for that purpose. This is not legal advice.

D := the dataset is CC BY 4.0     R := replies are teacher outputs
T := the teacher provider's terms bind a reuser of the replies
1. D         [project dataset license; this card]
2. R         [synthetic agent generation and ingestion described above]
3. R -> T    [teacher-terms caveat inherited from cards/DATASET_CARD.md; the card cannot waive terms]
∴ the dataset license does not remove the need to check the teacher providers' terms.