TSM4

28,477 synthetic English chats (user + assistant) for supervised fine-tuning, each with generation metadata. Mostly short: 1 to 3 exchanges.

Files and format

• data/data.parquet: Parquet, ZSTD, ~26 MB, 6 row groups.
• Columns: id (unique string), messages (JSON string: list of {role, content}, optional leading system), meta (JSON string, 25 fields, schema v2).
• Integrity scan: no duplicate or missing ids, no invalid metadata.

import json
from datasets import load_dataset
row = load_dataset("dragosMediv/TSM4", split="train")[0]
messages, meta = json.loads(row["messages"]), json.loads(row["meta"])

What it contains

meta.shape	Records	Share
one_turn	13,755	48.3%
two_turn	9,353	32.8%
three_turn	2,638	9.3%
short / medium / long (older, up to 16 turns)	2,731	9.6%

• Topics: 47 domains (Python, everyday life, casual chat, ML/AI, earth and space, philosophy, sports, ...), 404 subtopics, 41 task types (writing, factual questions, troubleshooting, recommendations, quizzes, comparisons, brainstorming, calculations, explanations, advice, planning, vague and trick questions).
• Style controls: greeting start none 86% / with request 11.5% / greeting only 2.3%; system prompt in 20.1%; answer length short 8,000 / medium 10,223 / long 4,533 and format light 10,112 / prose 7,980 / rich 4,664 (both null when a system prompt sets the style).
• Personas: ~11k distinct, 34 job labels, with varied expertise, tone and typing style.
• Size: 14.5M completion tokens; per record mean 510, median 299, max 11,364.
• Meta fields: identity (schema, harness, idx, seed_base), structure (shape, turns, greeting_mode, farewell, directives, ...), topic (domain, subtopic, task_type), persona, style, provenance (model, gen_seconds, completion_tokens) and qc. gen_seconds is wall time under heavy concurrency, not a benchmark.

How it was generated

Each record starts from a seeded random spec: domain and subtopic, task type, situation, persona (job, expertise, tone, typing style), greeting and farewell options, an optional system prompt (about 20 in the bank), and a preferred answer length and format. Multi-turn data adds moves like follow-ups, pushback, user mistakes, topic switches, distractions and format changes. Hidden assistant guidelines (be accurate, admit uncertainty, no filler openers, follow requested format) were used during generation but are not stored. Runs used two NVIDIA T4s.

meta.harness	Records	Model	Method
v4-oneshot	24,530	Qwen3-30B-A3B-Instruct-2507 (AWQ 4-bit, vLLM)	whole 1-3 exchange chat written in one call from a spec, shared cached system prompt
v2-vllm	1,745	same	turn by turn: simulated-user call plus assistant call per turn
v2-vllm2	856	same	like v2-vllm, trimmed prompts
v2	1,346	Qwen3.6-35B-A3B	earlier llama.cpp pipeline, turn by turn

Filtering: drops truncated output, broken turn order, CJK characters in English chats, generator-identity mentions (Qwen/Alibaba), unfilled placeholders, claimed-but-missing pasted content, repeated persona mentions, and leaked reasoning tags. Filler openers ("Great question!") are removed and stock closers mostly removed. v4-oneshot drops exact duplicate first messages within a run; there is no global dedup.

Quality

Only 172 records (0.6%) carry qc scores (1-5): user realism 4.84, assistant quality 4.33, format fit 4.94, 0 broken, 63 with flagged suspect facts. This is not a whole-dataset evaluation.

Limitations

• Synthetic: facts, code and numbers are not verified.
• English only; 95% from one model, so phrasing repeats (e.g. "Sure, here's a...", "You're absolutely right to..." in turn-by-turn data).
• Users are LLM-simulated, so more fluent and cooperative than real users.
• Few long conversations. Pipelines are mixed; filter by meta.harness.