persistent-fm/ctms-multitask-sft-v3

CTMS Multi-Task SFT — V3 (uppercase-Snowflake)

Supervised fine-tuning corpus for a Clinical Trial Management System (CTMS) analytics assistant, spanning 7 tasks over a 122-table CTMS schema. This is the V3 build: all SQL uses unquoted identifiers that resolve against the uppercase-identifier Snowflake schema DUMMY_FORTREA_AI_MODEL.FORTREA_AI_MODEL_V3_CAP.

Data is fully synthetic (generated from a CTMS data generator). It contains no real patient, investigator, or trial data.

Contents — 21,019 examples

task	examples	description
generate_sql	11,514	natural-language question → Snowflake SQL (Vanna-style prompt: schema context + DDL + guidelines)
sql_correction	2,000	broken SQL + execution error → corrected SQL (agentic-retry & regeneration subtasks)
trend_summarizer	2,000	conversation → trend question extraction + pointwise trend hints
summarizer	2,000	conversation → SQL-relevant question extraction + SQL hints
chart	1,505	result dataframe + goal → Plotly chart code (<think> + Python)
classifier	1,000	route a follow-up to SQL / PROFILING / CHART
insight	1,000	result dataframe → grounded natural-language insight
Total	21,019

Splits (80/10/10): train 16,817 · validation 2,101 · test 2,101. Seed-grouped (leakage-safe: examples sharing a source query move together), task-stratified, with verified-only evaluation.

Format

Chat-style, assistant-only loss:

{"messages": [{"role": "system", "content": "..."},
              {"role": "user", "content": "..."},
              {"role": "assistant", "content": "..."}],
 "meta": {"task": "generate_sql", "schema_variant": "v3_cap_uppercase", ...}}

SQL dialect / V3 note

SQL targets Snowflake with standard uppercase identifiers. Identifiers are written unquoted (e.g. SELECT s.study_code FROM study s) so Snowflake folds them to the stored uppercase names (STUDY, STUDY_CODE). A previous V2 build used quoted-lowercase identifiers ("study"."study_code") for a lowercase schema; this V3 build is the port to the uppercase schema.

Verification: every SQL-bearing answer — all 11,514 generate_sql answers and all 2,000 sql_correction corrected answers (13,514 total) — was executed live on FORTREA_AI_MODEL_V3_CAP and returns without error. DuckDB→Snowflake dialect issues were repaired (FIRST/LAST(x ORDER BY y) → MIN_BY/MAX_BY, INTERVAL n DAY → DATEADD, reserved word user → "USER", digit-leading aliases kept quoted). Static check confirms zero quoted-lowercase identifiers remain in any SQL.

Usage

from datasets import load_dataset
ds = load_dataset("persistent-fm/ctms-multitask-sft-v3")
print(ds)                      # train / validation / test
print(ds["train"][0]["messages"])

mixture.json provides suggested per-task sampling temperature weights (√-scaled) to prevent generate_sql from dominating a single multi-task model under uniform sampling.

Provenance & caveats

• Synthetic CTMS data; a hand-authored + execution-verified corpus.
• Prompts for generate_sql/sql_correction follow a Vanna-style envelope (schema context + DDL + response guidelines); chart uses a canonical Plotly-code prompt.
• Intended for SFT of a domain assistant that answers CTMS analytics questions and generates Snowflake SQL against the uppercase V3 schema.