license: apache-2.0
language: [en]
size_categories: [1K<n<10K]
tags: [sme, business-operations, document-management, data-management,
instruction-tuning, sft, qwen3.8, reasoning, thinking-model, agentic, tool-calling]
pretty_name: SME-SFT-DocData (Qwen3.8-27B)

SME-SFT-DocData (Qwen3.8-27B Native) — v2.0 production (9,342)

Agentic document + data management SFT dataset for Qwen3.8-27B
(abliterated or base). Teaches the model to manage business documents and data
for small and medium enterprises: search a document corpus, read and summarize
docs, browse folders, and run real aggregations / filters / stats over CSV tables —
then ground every answer in what the tool actually returned. Companion to
hotdogs/cyber-sft-agent-qwen38 (the security tool layer); this v2.0 adds
RAG rows (passive retrieval) and advanced rows (error recovery, multi-doc
synthesis, temporal reasoning, conflict resolution) for production-grade robustness.

v2.0 composition

• 9,342 rows — train 7,799 / valid 1,028 / test 515 (stratified)
• 2,520 tool rows — 11 tools, 16 chain archetypes, every tool output computed from a deterministic per-seed world (a fictional SME's real folder tree + CSVs) so the evidence is consistent, not hallucinated
• 5,880 QA rows — pure document/data-management QA across 5 families (define / process / why / compare / situation), parameterized over 6 SME industries
• 192 RAG rows — passive retrieval: context provided inline (no tool calls), teaches model to reason from retrieved context directly
• 750 advanced rows — error recovery (200), multi-doc synthesis (300), temporal reasoning (150), conflict resolution (100)

11 tools

doc_search · doc_read · doc_sections · doc_summarize · doc_metadata · folder_list · csv_head · csv_aggregate · csv_filter · csv_stats · doc_compare

Document tools: doc_search, doc_read, doc_sections, doc_summarize, doc_metadata, folder_list.
Data tools: csv_head, csv_aggregate, csv_filter, csv_stats. Cross-doc: doc_compare.

Chain archetypes (2,520 tool rows)

• Single-tool — one well-chosen call for a focused ask (find a doc, read it, roll up a column, filter a table).
• D1-D9 multi-tool — the real SME workflows: search→read, search→summarize, head→aggregate, filter→stats, folder→metadata, sections→read, two-aggregate compare, doc compare, and D9 cross-check (a report makes a claim; the model verifies it against the raw CSV and says whether they agree).
• N1-N3 negative evidence — anti-hallucination: an honest "not found" (N1), an empty filter that is correct because the threshold is above the max (N2), and a doc that exists but doesn't answer the question (N3).

Quality gates (all passing)

Byte-exact: all 9,342 rows re-render through the real Qwen3.8-27B chat
template and verified 9,342/9,342 = 100.0%, 0 render errors.
Token length measured with the real tokenizer (Qwen/Qwen3.5-0.8B (vocab 248044)).

Format (the important part)

Rendered through the real Qwen3.8-27B template and verified byte-exact.
Format facts (verified via byte ordinals):

Row schema

• id — stable row id
• kind — tool (agentic chain), qa (pure QA), rag (passive retrieval), error_recovery, multi_doc, temporal, conflict
• tags — e.g. search, data, aggregate, multi, negative, qa, rag, error, temporal
• tools — function schemas exposed to the model (used tools + optional distractors; empty for RAG/advanced rows)
• messages — template-native list, roles system/user/assistant/tool; thinking in reasoning_content, calls in tool_calls: [{name, arguments:{dict}}]
• meta — n_tool_calls, n_turns, chain, world, industry
• text — the full pre-rendered template string (use for packing / single-text SFT)

Splits (stratified)

Usage

A. messages[] (trainer calls apply_chat_template)

from datasets import load_dataset
ds = load_dataset("hotdogs/sme-sft-docdata-qwen38", split="train")
row = ds[0]
messages, tools = row["messages"], row["tools"]

B. text (packing / single-text SFT, no template needed — recommended)

text = row["text"]   # already byte-exact through the Qwen3.8-27B template

Recommended training config (Qwen3.8-27B, 4× RTX 3060 12GB)

lora:
  r: 64
  alpha: 128
  target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
training:
  seqlen: 4096        # covers 100% (real max 3,729 tok)
  batch_size: 4
  gradient_accumulation: 2
  learning_rate: 2e-4
  epochs: 3

Provenance

Generated by the sme-sft pipeline (Hermes Agent). Tool rows are produced from
deterministic per-seed worlds — each world is a fictional SME with a real folder
tree of documents (contracts, invoices, reports, HR policies, meeting minutes, product
sheets) plus CSV tables (sales, invoices, employees, expenses, customers); every tool
output is computed from that world, so a summary matches the doc and an aggregate
matches the CSV. QA rows are drawn from a combinatorial generator over 6 SME
industries and 5 question families. Date anchor: 2026-08-20. Format ground truth:
byte-ordinal analysis of the real template, validated 100% against
hotdogs/cyber-sft-agent-qwen38.