Add arXiv citation and paper link (arXiv:2605.25038)

README.md

@@ -1,6 +1,8 @@
 ---
 pretty_name: "TRACE — Taxonomy-Referenced ABA Clinical Examples"
 license: cc-by-nc-4.0
+arxiv:
+  - 2605.25038
 language:
   - en
 size_categories:
@@ -32,13 +34,15 @@ configs:
 # TRACE Dataset Card (v1)
 
 **Name:** **TRACE** — **T**axonomy-**R**eferenced **A**BA **C**linical **E**xamples
-**Version:** v1 (dataset hash pinned to commit `945248d`)
-**Date:** 2026-04-24
+**Version:** v1.0.0
+**Date:** 2026-04-25
 **Primary language:** English
-**License:** Synthetic data, released under [TBD — recommend CC BY-NC 4.0 for research use; no clinical deployment].
+**License (data):** CC BY-NC 4.0 · **License (code):** MIT
 **Total examples:** 2,999
 **Tasks:** 2 — (1) ABA teaching program generation, (2) behavioral session interpretation.
-**Maintained by:** Festus Kahunla (Drexel University).
+**Author:** Festus Kahunla (Drexel University).
+**Publisher / maintained by:** [Pombo Labs](https://github.com/Pombo-Labs).
+**Repository:** https://github.com/Pombo-Labs/TRACE
 
 ---
 
@@ -51,100 +55,9 @@ TRACE is a **synthetic instruction-tuning dataset** for two clinical tasks in Ap
 
 The dataset was produced by a **taxonomy-driven generator** whose controlled vocabulary is grounded in the canonical ABA literature (Cooper, Heron, & Heward 2020; VB-MAPP; AFLS; key JABA papers). Every example carries full **provenance metadata** — the exact taxonomy cells that were sampled to produce it. Clinical accuracy was iterated via practitioner-in-the-loop ad-hoc review.
 
-**Intended use.** Fine-tune small language models (4-bit Gemma 4 E2B with QLoRA) for on-device decision support **drafting**. Not for autonomous clinical decisions. A Board Certified Behavior Analyst (BCBA) remains responsible for every program and every interpretation.
+**Intended use.** Research. TRACE is designed for fine-tuning small language models (e.g., 4-bit Gemma 4 E2B with QLoRA) on ABA-flavored instruction-following, as a substrate for research into clinical-NLP data pipelines, taxonomy-driven synthetic generation, and small-LM evaluation.
 
-**Not for:** training on real client data; replacing a BCBA; writing final Behavior Intervention Plans without clinical review; clinical use without facility approval.
-
-### 1.1 Inputs, outputs, and evaluation
-
-TRACE is a supervised fine-tuning corpus. For both tasks, the **model input** is `messages.user`, the **supervised target** is `messages.assistant`, and `meta.gold_labels` holds extractable categorical labels used for automated scoring. The system prompt (`messages.system`) is identical across all examples within a task and is not part of the input the model has to interpret.
-
-#### Task 1 — Teaching program generation
-
-| | |
-|---|---|
-| **Input** | A BCBA-style request specifying skill target, curriculum reference, learner profile, and current mastery state. |
-| **Output** | A structured Markdown teaching program with sections for overview, end goal / trial procedure, prompt strategy, error correction, reinforcement, mastery criteria, data collection, and generalization. |
-| **Evaluation** | (a) Categorical accuracy on `meta.gold_labels.method` (`dtt` / `net` / `task_analysis`) and `meta.gold_labels.program_type` (Task Analysis only: `independence` / `toleration`); (b) BERTScore F1 between the generated assistant message and the gold assistant message. |
-
-**Worked example (trimmed):**
-
-```
-INPUT (messages.user):
-  Skill Target: tolerating tooth brushing by caregiver
-  Curriculum: AFLS Basic Living Skills
-  Learner: School-Age Learner
-  Mastery Status: Generalization phase
-  Please include the toleration end goal, shaping progression,
-  antecedent strategies, reinforcement plan, refusal / safety
-  procedures, and mastery criteria.
-
-OUTPUT (messages.assistant, abbreviated):
-  ## Program Overview
-  - Target skill: tolerating tooth brushing by caregiver
-  - Approach: Toleration / systematic desensitization
-  ## Shaping Progression
-  1. 1-second toothbrush contact with teeth
-  2. 3 seconds of brushing
-  ...
-  ## Reinforcement
-  Deliver high-magnitude, high-preference reinforcement contingent
-  on successful toleration of the target duration.
-  ## Mastery Criteria
-  End-goal activity tolerated with 2 different caregivers across
-  2 consecutive weeks.
-
-GOLD LABELS (meta.gold_labels):
-  method:          "task_analysis"
-  program_type:    "toleration"
-  domain:          "AFLS.basic_living"
-  learner_profile: "school_age"
-  mastery_state:   "generalization"
-```
-
-#### Task 2 — Behavioral session interpretation
-
-| | |
-|---|---|
-| **Input** | A multi-session behavioral log (5–12 sessions): learner profile, acceleration programs with per-session accuracy / latency / prompt-distribution data, deceleration targets with measurement traces, and (in ~30% of logs) ABC entries and IOA agreement values. |
-| **Output** | A structured Markdown interpretation with sections for clinical concerns, pattern classification, behavior function hypothesis, programming recommendations (antecedent / replacement / consequence / crisis), escalation level, confidence, and data-supported rationale. |
-| **Evaluation** | (a) Categorical accuracy on `pattern_class` (12-way), `behavior_functions` (5-way per behavior), `escalation_level` (1–5; exact-match accuracy and MAE), `confidence` (3-way), and `crisis_plan_required` (binary F1); (b) BERTScore F1 against the gold assistant message. |
-
-**Worked example (trimmed):**
-
-```
-INPUT (messages.user, abbreviated):
-  LEARNER PROFILE: Early Learner, 4 acceleration programs,
-                   1 deceleration target
-  Session 1 (2026-06-10): mand 4/10 (38%); listener 5/12 (42%); ...
-  Session 2 (2026-06-12): mand 6/14 (44%); listener 4/9 (47%); ...
-  ...
-  Session 8 (2026-06-24): mand 9/12 (72%); listener 14/15 (92%); ...
-  Vocal stereotypy: ranged freq 0–1 across sessions
-
-OUTPUT (messages.assistant, abbreviated):
-  ## Clinical Concerns
-  Steady improvement: accuracy rose from 38% to 72% across 8 sessions.
-  ## Pattern Classification
-  mastery_progression
-  ## Behavior Function Hypothesis
-  Vocal stereotypy: automatic
-  ## Escalation Level
-  1 — Continue monitoring
-  ## Confidence
-  moderate
-
-GOLD LABELS (meta.gold_labels):
-  pattern_class:        "mastery_progression"
-  behavior_functions:   { "Vocal stereotypy": "automatic" }
-  escalation_level:     1
-  confidence:           "moderate"
-  crisis_plan_required: false
-```
-
-#### Why two scoring channels
-
-The assistant message is **free-form clinical prose** but the underlying clinical decision is **categorical** (which method, which pattern, what escalation level). We score both layers because a generation that hits the labels but produces unsafe prose is a failure, and so is eloquent prose with the wrong pattern classification. The categorical labels capture clinical correctness; BERTScore captures prose alignment with the taxonomy-grounded reference.
+**Not for:** autonomous clinical decisions; training on or combining with real client data; medical diagnosis; legal or insurance documentation. TRACE has not been clinically validated and is not a clinical tool. See section 6 for the responsibility disclaimer.
 
 ---
 
@@ -155,9 +68,9 @@ The assistant message is **free-form clinical prose** but the underlying clinica
 | Split | Examples | Fraction | Purpose |
 |---|---:|---:|---|
 | `train.jsonl` | 2,549 | 85.0% | LoRA fine-tuning |
-| `valid.jsonl` | 149 | 5.0% | Periodic validation loss during training |
-| `test.jsonl` | 281 | 9.4% | Held-out evaluation (headline metrics) |
-| `sanity.jsonl` | 20 | 0.7% | Training smoke-test (tiny stratified subset) |
+| `valid.jsonl` | 149 | 4.97% | Periodic validation loss during training |
+| `test.jsonl` | 281 | 9.37% | Held-out evaluation (headline metrics) |
+| `sanity.jsonl` | 20 | 0.67% | Training smoke-test (tiny stratified subset) |
 | **Total** | **2,999** | 100% | |
 
 Splits are **stratified by task × category** (method for teaching programs; pattern_class for session interpretation) so each split mirrors the corpus distribution. The test set is the full curation pool minus a 20-example stratified sanity carveout.
@@ -168,7 +81,7 @@ Splits are **stratified by task × category** (method for teaching programs; pat
 |---|---:|---|---|
 | DTT | 800 | teaching_program | VB-MAPP (array-based discrete-response domains) |
 | NET | 500 | teaching_program | VB-MAPP (mand, social, spontaneous vocal, intraverbal) |
-| Task Analysis | 499 | teaching_program | AFLS (basic_living, home, community, vocational, independent_living) |
+| Task Analysis | 500 | teaching_program | AFLS (basic_living, home, community, vocational, independent_living) |
 | Session Interpretation | 1,200 | session_interpretation | 12 clinical trajectory patterns |
 
 ### 2.3 DTT skill-domain distribution (800 total)
@@ -196,7 +109,7 @@ Sampled across VB-MAPP Levels 1 (≈45%), 2 (≈40%), and 3 (≈15%).
 
 Each NET program carries a Motivating Operation arrangement matched to the skill (deprivation, missing-item, break opportunity, completion opportunity, bathroom opportunity, peer presence, reciprocal conversation, routine lead-in) and is embedded in a natural context (snack, free play, transition, arrival, etc.).
 
-### 2.5 Task Analysis distribution (499 total)
+### 2.5 Task Analysis distribution (500 total)
 
 | AFLS module | Count | | Program type | Count |
 |---|---:|---|---|---:|
@@ -354,27 +267,11 @@ For each example:
 - **Teaching methods** — DTT (Lovaas 1987; Smith 2001); NET (Hart & Risley 1975 in CHH Ch. 18); Task Analysis / chaining (CHH Ch. 20).
 - **Operational definitions of target behaviors** — Cooper/Heron/Heward Ch. 3, 27; key JABA papers (Iwata et al. 1994 for SIB and functional analysis; Carr & Durand 1985 for FCT; Hanley, Iwata, & McCord 2003 for FBA).
 - **Session patterns** — derived from CHH Ch. 6–7 (analyzing behavior change) and Stokes & Baer 1977 (generalization).
-- **Crisis plans** — BACB Ethics Code (2020) §3.05; ABAI Position Statement on Restraint and Seclusion (2010). Physical-intervention procedures are left vague on purpose because they vary by training program (Safety-Care, CPI, PMT, TCI) and jurisdiction; the dataset emphasizes verbal de-escalation, environmental safety, BIP authorization, and contraindications.
+- **Crisis plans** — BACB Ethics Code (2020) section 3.05; ABAI Position Statement on Restraint and Seclusion (2010). Physical-intervention procedures are left vague on purpose because they vary by training program (Safety-Care, CPI, PMT, TCI) and jurisdiction; the dataset emphasizes verbal de-escalation, environmental safety, BIP authorization, and contraindications.
 
 ### 4.4 Clinical-accuracy pipeline
 
-Initial generation produced the structural skeleton. The corpus was then iterated via **practitioner-in-the-loop review**: a full-text render of every held-out candidate was browsed by a reviewer with ABA practitioner exposure, flags were raised for clinically inaccurate topographies, and each flag triggered a taxonomy or rendering fix plus a regeneration and a commit. Examples of iterations shipped:
-
-| Issue flagged | Fix |
-|---|---|
-| "Context: R/D" tag leaking a specific facility's tool | Removed from all generators and templates |
-| "Raises hand to request" wrong topography for mand | Replaced with AAC, vocal, or full-sentence requesting; added Tolerating Denied Access as a replacement-behavior target |
-| "Whining during trials" in frustration indicators | Removed; replaced with property destruction, aggression, peer aggression, SIB onset |
-| Pica measured as simple frequency | Split into attempts / successful / unsuccessful |
-| Toleration programs missing entirely | Added as a distinct program type for hygiene routines |
-| Crisis plans too generic | Rewritten with tiered framework, BIP-authorization requirement, contraindications, BACB Ethics citations |
-| `maintenance` mastery state (post-mastery retention) appeared as a current teaching state | Removed; maintenance is retained only as a post-mastery probe concept in generalization plans |
-| Only one dishwashing variant in Home Skills | Added by-hand, operating dishwasher, unloading dishwasher |
-| Token-economy backups abstract | Made concrete for children (candy, cookies, leisure, screen time, sensory item) |
-| Fecal smearing and toileting accidents absent | Added as target behaviors with attempts/intercepted/completed and urine/BM in-toilet/accidents rendering |
-| Bathroom requesting absent from mand curriculum | Added to NET L1 mands with bathroom_opportunity MO |
-
-Each iteration is one commit — the provenance chain is preserved end-to-end.
+Initial generation produced the structural skeleton. The corpus was then iterated via **practitioner-in-the-loop review** — a full-text render of the held-out candidate pool was browsed by a reviewer with ABA practitioner exposure, and each flagged clinical inaccuracy was traced to the responsible taxonomy cell and fixed with a single targeted edit plus a full regeneration. Because every example records its sampling provenance, a single cell-level edit propagates to every example that sampled the affected cells — so flagging one example systematically corrects a class of examples.
 
 ---
 
@@ -382,11 +279,11 @@ Each iteration is one commit — the provenance chain is preserved end-to-end.
 
 ### 5.1 Direct use
 - Instruction-tuning a small language model (recommended: Gemma 4 E2B 4-bit with QLoRA) to draft ABA teaching programs and interpret session logs.
-- Evaluation of task-specific competencies using the held-out `test.jsonl` (279 examples).
+- Evaluation of task-specific competencies using the held-out `test.jsonl` (281 examples).
 - Research on taxonomy-driven synthetic data generation for clinical decision support.
 
 ### 5.2 Downstream use
-- Pilot **drafting assistants** for BCBAs, always requiring clinician review before use with learners.
+- Research on small-LM drafting assistants in structured clinical-documentation domains.
 - Comparison baselines for future ABA-specific LLM work (Kumar et al. 2024 "Personalized-ABA" is the closest direct predecessor; the present dataset extends to structured program generation and session-log interpretation).
 
 ### 5.3 Out-of-scope (do NOT use for)
@@ -404,10 +301,10 @@ Each iteration is one commit — the provenance chain is preserved end-to-end.
 Every example is synthetic. No real client data, no real session notes, no real identifiers were used at any step. Learner references use synthetic IDs (`SYN-####`); dates fall in the range 2026-01-01 to 2026-12-31.
 
 ### 6.2 Clinical-risk framing
-The dataset is framed — in system prompts, in the dataset card, and in the paper — as producing **drafts for clinician review**. Assistant responses are structured so that a BCBA can quickly see the method, the stimulus arrangement, the prompt hierarchy, the reinforcement plan, the error-correction procedure, the mastery criterion, and the generalization plan; approve what is clinically appropriate; and revise what is not. Session-interpretation responses surface a confidence level (high / moderate / low) and an escalation level (1–4) explicitly so the clinician can calibrate action.
+The dataset is designed around a **draft-and-review** authoring pattern. Assistant responses are structured so that a reviewer can quickly see the method, the stimulus arrangement, the prompt hierarchy, the reinforcement plan, the error-correction procedure, the mastery criterion, and the generalization plan — each as a distinct, scannable section. Session-interpretation responses surface a confidence level (high / moderate / low) and an escalation level (1–4) as structured fields. These are design choices that support auditability; they are not clinical advice, and TRACE's responsibility disclaimer applies.
 
 ### 6.3 Crisis-plan sensitivity
-Crisis plans were written against the ABAI 2010 Position Statement on Restraint and Seclusion and BACB Ethics Code 2020 §3.05. The dataset references facility crisis-prevention frameworks (Safety-Care, CPI, PMT, TCI) only as examples and **deliberately avoids specifying restraint procedures** because those procedures are (a) jurisdiction-dependent, (b) training-certification-gated, and (c) learner-specific (many learners have contraindications). The dataset embeds explicit text in every crisis-plan bullet that physical intervention is used only when specifically authorized in the learner's BIP and only by staff currently certified in the facility's training program.
+Crisis plans were written against the ABAI 2010 Position Statement on Restraint and Seclusion and BACB Ethics Code 2020 section 3.05. The dataset references facility crisis-prevention frameworks (Safety-Care, CPI, PMT, TCI) only as examples and **deliberately avoids specifying restraint procedures** because those procedures are (a) jurisdiction-dependent, (b) training-certification-gated, and (c) learner-specific (many learners have contraindications). The dataset embeds explicit text in every crisis-plan bullet that physical intervention is used only when specifically authorized in the learner's BIP and only by staff currently certified in the facility's training program.
 
 ### 6.4 Population representation
 Learner profiles are intentionally abstract (early / school-age / adolescent / adult). The dataset does not encode demographic categories (race, socioeconomic status, gender identity) and does not attempt to characterize clinical presentations by such categories. This is a deliberate choice for a first release; future versions may add representation if grounded in published demographic work.
@@ -418,10 +315,12 @@ Learner profiles are intentionally abstract (early / school-age / adolescent / a
 - **VB-MAPP + AFLS only.** Other curricula (ABLLS-R, Essential for Living, PEAK) are not covered. Practitioners using those curricula should adapt.
 - **No longitudinal data.** Sessions within a log are temporally ordered but the pipeline does not model real continuity over months or years.
 - **Toleration covers hygiene only.** Other toleration programs (e.g., wearing glasses, riding in a car seat) are not represented.
-- **Toilet-training acquisition is out of scope.** Accidents are tracked in session logs and bathroom-requesting is taught as a NET mand, but the full Azrin & Foxx 1971 rapid toilet-training acquisition protocol is not included as a task-analysis program. This is a planned v2 addition.
+- **Toilet-training acquisition is out of scope.** Accidents are tracked in session logs and bathroom-requesting is taught as a NET mand, but the full Azrin & Foxx 1971 rapid toilet-training acquisition protocol is not included as a task-analysis program.
 
 ### 6.6 License
-TBD — recommend CC BY-NC 4.0 (research use, non-commercial) pending the hackathon submission requirements. Clinical use of the dataset or any model fine-tuned on it requires facility approval and BCBA oversight.
+**Data:** CC BY-NC 4.0 — research and non-commercial use with attribution. **Code:** MIT.
+
+**Responsibility.** TRACE is a research artifact. It is not a clinical tool, has not been clinically validated, and carries no clinical endorsement. Anyone who chooses to deploy TRACE — or any model derived from it — in a clinical setting does so entirely at their own responsibility and under their facility's own oversight. The authors and Pombo Labs make no representation of clinical suitability and accept no liability for clinical outcomes.
 
 ---
 
@@ -433,28 +332,52 @@ The corpus is regenerated deterministically from:
 - `src/generators/*.py` (generator code)
 
 ```bash
-uv run python src/generate.py --all                  # regenerate the 2999-example corpus
+uv run python src/generate.py --all                  # regenerate 3000 examples
 uv run python src/split_data.py                      # stratified split
 uv run python src/prepare_curation.py                # browseable review.md
 uv run python src/compile_curation.py                # test.jsonl + sanity.jsonl
 ```
 
-The repo commit pinned for this dataset version is `945248d`.
+The dataset version is `v1.0.0`; the matching git tag pins the exact configs and generator code that produced the published JSONL splits.
 
 ---
 
 ## 8. Citation
 
-Please cite as:
+**Paper:**
+
+```bibtex
+@misc{kahunla2026tracetaxonomygroundedsyntheticdataset,
+  title         = {TRACE: A taxonomy-grounded synthetic dataset for teaching-program generation and session interpretation in Applied Behavior Analysis},
+  author        = {Festus Kahunla},
+  year          = {2026},
+  eprint        = {2605.25038},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.CL},
+  url           = {https://arxiv.org/abs/2605.25038}
+}
+```
+
+**Dataset:**
+
+```bibtex
+@dataset{trace_2026,
+  title     = {TRACE: Taxonomy-Referenced ABA Clinical Examples},
+  author    = {Kahunla, Festus},
+  year      = {2026},
+  publisher = {Pombo Labs},
+  url       = {https://github.com/Pombo-Labs/TRACE},
+  note      = {Version 1.0.0}
+}
+```
 
-> Kahunla, F. (2026). *TRACE: Taxonomy-Grounded Synthetic Data for Teaching Program Generation and Session Interpretation in Applied Behavior Analysis.* Drexel University CS 614 final project / Gemma 4 Good Hackathon submission.
+Machine-readable metadata: `CITATION.cff`.
 
 ---
 
 ## 9. Appendices
 
-- **Datasheet** (Gebru et al. 2021 format): `docs/dataset/datasheet.md`
-- **Data statement** (Bender & Friedman 2018 format): `docs/dataset/data-statement.md`
-- **Taxonomy reference** (operational definitions + citations): `docs/dataset/taxonomy-v1.md`
-- **Schema reference** (wire format + slot specifications): `docs/dataset/schema-v1.md`
-- **Research paper draft:** `docs/paper/paper-draft.md`
+- **Datasheet** (Gebru et al. 2021 format): `datasheet.md`
+- **Data statement** (Bender & Friedman 2018 format): `data-statement.md`
+- **Taxonomy reference** (operational definitions + citations): `taxonomy-v1.md`
+- **Schema reference** (wire format + slot specifications): `schema-v1.md`