Nanthasit/sakthai-combined-v6

SakThai Combined v6 — Improved

Expanded tool-calling dataset for SakThai-Agent fine-tuning — with Hugging Face Hub tools, Growth Cycle integration, multi-turn conversations, and 10 new Hermes agent tool schemas.

Dataset Structure

File	Records
data/train.jsonl	908
data/test.jsonl	113 (54 with tool calls)

What's New in v6.1 (2026-07-13)

180 hand-crafted records added (scripts/generate_v61_*.py), targeting behavioral gaps:

Addition	Count	Why
No-tool negatives — tools present in schema but the correct behavior is a direct answer	~100	Train had only 1/753 direct-answer records; model would learn to always call a tool
Thai coverage — Thai user → tool call → Thai answer, plus Thai direct answers	88 records now contain Thai (train 8%, test 11%)	Dataset was 100% English despite the SakThai target
Rare-tool bulking — session_search, cronjob, hf_upload_file, web_search, search_web, capture_lesson, run_agent_loop	+32	All tools now have ≥5 examples
clarify examples — ambiguous requests answered with a clarifying tool call	+5	Teaches asking instead of guessing
Param descriptions for the 29 most-used tools, applied to every record	1,300+ params filled	Better argument accuracy

Quality Pass (v6 Cleaned, 2026-07-13)

Full-dataset audit + repair before training (scripts/fix_quality.py):

Fix	Count
Tool called but missing from record's tools list → schema injected	110
Legacy-format calls: missing id added	160
arguments as dict → JSON string (OpenAI format)	135
tool_calls as dict → wrapped in list	7
Tool schema description filled from sibling schemas	581
Parameter descriptions filled from sibling schemas	308
Curly apostrophes in system prompts normalized	163
Tool-calling records moved train → test (34 distinct tools)	40

Post-fix audit: 0 structural errors, 0 orphan tool_call IDs, 0 tools called outside schema, 0 train/test leakage (exact or first-prompt). Note: 160 single-turn records intentionally end at the assistant tool call (no tool response) — the training target is call emission.

What's New in v6 Improved

Improvement	Before	After	Δ
Training examples	663	789	+126 (+19%)
Tool schemas	71	81	+10
Tool calls	932	1,112	+180 (+19%)
Multi-turn (3+ user turns)	14	22	+8 (+57%)

10 New Hermes Tool Schemas

search_files, patch_file, skill_manage, skill_view, todo, memory_store, browser_navigate, browser_click, clarify, vision_analyze

Quality

• Multi-turn conversations with proper user → assistant → tool → user → ... turn structure
• Rare schemas bulked — all 11 under-represented tools now have 5-10 examples
• Zero short responses — every assistant message is a natural, complete sentence
• Trust-passed — 0 structural errors, 0 orphan IDs in new data

New Workflow Examples

Research (6 turns)

User → assess_energy + hf_search_models
User → hf_get_model_card
User → hf_run_inference
User → supermemory_store
User → log_transition + close_cycle

Skill Building (4 turns)

User → skill_manage(create)
User → skill_view
User → todo(create)

Version History

Version	Records	Tools	Features
v5	659	53	Generic tool-calling, parallel calls
v6	663	71	+ HF Hub + Growth Cycle + real tools
v6 Improved	789	81	+ multi-turn + new schemas + rare schema bulking
v6 Cleaned	753 + 88 test	81	format normalization + schema injection + tool-calling test split
v6.1	908 + 113 test	81	no-tool negatives + Thai coverage + rare-tool bulking + clarify + param descriptions