กำเมือง · Lanna-3L Codenames

❖ ❖ ❖

กำเมือง

Lanna Mnemonics for LannaBench Project

Every module, package, and command in the Lanna-3L benchmark carries an authentic Northern Thai name — drawn from traditional craft, ceremony, and community life of the ancient Lanna Kingdom.

ข่วง · Layer A · Data คัวไฟ · Layer B · Adaptation จ่อฟ้า · Layer C · Evaluation อูบ · Utilities

orchestrated by ฟ้อน · $ lanna-fon run --pipeline paper1_full.yaml

28Codenames

3Layers

explore

Layer A · lanna_khuang/

ข่วง Khuang · Courtyard Data Foundation

น้ำต้นหูกตอกรักตุง

Layer B · lanna_kuafai/

คัวไฟ Kua Fai · Kitchen Adaptation & Inference

ขันโตกไหข้าวกัวะอั้งโล่

Layer C · lanna_jorfa/

จ่อฟ้า Jor Fa · Sky Finial Evaluation & Diagnostics

สาตุไตรภูมิแก้วป่องแตะใบลานฮอมปอยกาแล

Utilities · oob_meta/

อูบ Oob · Lacquerware Container Shared Utilities

ปู่จ๋านป่อหลวงข้าวตอกปี้ฮีตฟ้อน

❖

Layer A · lanna_khuang/ ข่วง Khuang · Open Courtyard "The open gathering space where community begins all important activity" $ lanna-khuang · formerly lannaforge · Data Foundation Toolkit

ข่วง

Khuang

"The open courtyard before a temple or palace"

Raw materials are laid out in the courtyard, organised, and prepared before any craft begins. Layer A takes the raw master Google Sheet and builds the JSONL corpus — everything downstream begins here.

$ lanna-khuang

น้ำต้นNam Ton · Water Carafe→ namton_builder.py · JSONL builder

formerly Anvil

click to expand ↗

หูกHuuk · Loom→ huuk_validator.py · schema validator

formerly Loom

click to expand ↗

ตอกTok · Bamboo Strips→ tok_splitter.py · partitioner

formerly Cleaver

click to expand ↗

รักRak · Lacquer→ rak_annotator.py · codebook applicator

formerly Quill

click to expand ↗

ตุงTung · Ritual Flag→ tung_definer.py · field definitions

formerly Compass

click to expand ↗

❖

Layer B · lanna_kuafai/ คัวไฟ Kua Fai · Traditional Kitchen "The heart of transformation — where fire and skill shape everything" $ lanna-kuafai · formerly lannasmith · Adaptation Toolkit

คัวไฟ

Kua Fai

"The traditional Lanna kitchen — spiritual and architectural heart of the home"

Not just a cooking room but the centre of transformation. Raw ingredients become meals; fire shapes everything. Layer B transforms raw models through zero-shot prompting, few-shot conditioning, and LoRA fine-tuning. The GPU is the fire.

$ lanna-kuafai

ขันโตกKhan Tok · Serving Bowl→ khantok_registry.py · 9-model registry

formerly Roster

click to expand ↗

ไหข้าวHai Khao · Rice Steamer→ haikhao_templater.py · prompt templating

formerly Stencil / Chalu

click to expand ↗

กัวะKua · Rice Cooling Tray→ kua_inference/ · inference runner

formerly Forgehammer

click to expand ↗

อั้งโล่Ang Lo · Charcoal Brazier→ anglo_finetuner.py · LoRA fine-tuner

formerly Bellows / Sin

click to expand ↗

❖

Layer C · lanna_jorfa/ จ่อฟ้า Jor Fa · Sky Finial "Reaching skyward — connecting the earthly generation to the heaven of perfect accuracy" $ lanna-jorfa · formerly lannascope · Diagnostic Toolkit

จ่อฟ้า

Jor Fa

"The elegant finial atop a Lanna viharn, often carved as a naga"

It reaches skyward, connecting earth to the highest realms — a measuring rod of sacred aspiration. Layer C runs the full LannaDiagnos stack: Triple-ChrF, Dialect Gap Score, LannaFacets slicing, human evaluation, and error analysis.

$ lanna-jorfa

① Automatic Diagnostics

สาตุSatu · Affirmation→ satu_robustness.py · 3P robustness

formerly Trident

click to expand ↗

ไตรภูมิTrai Phum · Three Realms→ traiphum_multiref.py · MultiRef scorer

formerly Triplet

click to expand ↗

แก้วKaew · Crystal→ kaew_aggregator.py · Triple-ChrF

formerly Prism

click to expand ↗

ป่องPong · Gap / Interval→ pong_gap.py · Dialect Gap Score G

formerly Caliper

click to expand ↗

แตะTae · Woven Bamboo Panel→ tae_slicer.py · LannaFacets slicer

formerly Lattice

click to expand ↗

② Human Diagnostics

ใบลานBai Lan · Palm-Leaf Manuscript→ bailan_form.py · human eval forms

formerly Slate

click to expand ↗

ฮอมปอยHom Poi · Communal Merit→ hompoi_agreement.py · Krippendorff α

formerly Concord

click to expand ↗

③ Error Analysis

กาแลKalae · Guardian Gable Finial→ kalae_error.py · error analysis

formerly Sieve / Hae

click to expand ↗

❖

Utilities · oob_meta/ อูบ Oob · Ceremonial Lacquerware Container "A hand-crafted lacquerware container that holds and preserves what is most sacred" formerly lannakit · Shared utilities across all layers

อูบ

Oob

"In Lanna and Tai Khoen culture, an อูบ is a hand-crafted lacquerware container decorated with intricate gold-leaf patterns"

Used to store Buddhist manuscripts, ritual betel sets, or offerings for monks — precious, sacred items that must be preserved intact. The อูบ_meta package holds all shared utilities: schema definitions, run manifests, seed manager, I/O bridge, and orchestration. Like a ceremonial container, it safeguards the rules that ensure reproducibility.

oob_meta/

ปู่จ๋านPu Jan · Lay Minister→ pujan_io.py · JSONL I/O bridge

formerly io.py

click to expand ↗

ป่อหลวงPo Luang · Village Headman→ poluang_orchestrate.py · pipeline leader

formerly orchestrate.py

click to expand ↗

ข้าวตอกKhao Tok · Sacred Popped Rice→ khaotok_seed.py · deterministic seed

formerly loomstone.py

click to expand ↗

ปี้Pee · Tag / Mark→ pee_manifest.py · run manifest

formerly manifest.py

click to expand ↗

ฮีตHeet · Customary Law→ heet_schema.py · Pydantic schemas

formerly schema.py

click to expand ↗

ฟ้อนFon · Traditional Dance→ $ lanna-fon run · pipeline orchestrator

formerly lanna run

click to expand ↗

❖

The Dance in Motion

ฟ้อน · lanna-fon

The lanna-fon command is the pipeline orchestrator — it reads a YAML pipeline file and runs every layer in sequence, passing manifest IDs between stages. Think of it as the village headman (ป่อหลวง) calling the performance: each tool plays its part in order, and the benchmark runs from raw sheet to final report in one graceful, uninterrupted dance. Below is a realistic terminal session showing the full Paper 1 pipeline.

northern-thai-llm · terminal

Layer A · ข่วง · Build the corpus

$lanna-khuang build --sheet lanna_master.xlsx --out data/

✓ namton_builder 1120 rows forged into JSONL ✓ tok_splitter test=200 · dev=100 · train=820 (stratified)

$lanna-khuang validate data/lannacore_dataset.jsonl

✓ huuk_validator 1120 rows comply with ฮีต · 0 violations

Layer B · คัวไฟ · Fire up the kitchen

$lanna-kuafai register

khantok_registry 9 models loaded (5 api · 4 fine-tunable)

$lanna-kuafai infer --model gpt4o --condition zero-shot --prompt-index 1

✓ haikhao_templater → kua_inference/api 200 items · 45.1s

$lanna-kuafai train --base typhoon2 --r 8 --alpha 16

✓ anglo_finetuner LoRA adapter infused · pee_manifest stamped

Layer C · จ่อฟ้า · Reach toward the sky

$lanna-jorfa score --hypotheses outputs/ --references test.jsonl

✓ traiphum_multiref → kaew_aggregator chrf_max=42.1 · chrf_avg=38.7 · chrf_diff=3.4

$lanna-jorfa gap --reports reports/

✓ pong_gap G = 8.3 (STD − NTD dialect gap)

$lanna-jorfa errors --reports reports/ --human human/responses/

✓ kalae_error 20 worst-k · 12 tone-flip · 8 CD-drop flagged

Full pipeline · ฟ้อน · One graceful dance

$lanna-fon run --pipeline pipelines/paper1_full.yaml

🕊️ Lanna Fon — orchestrating the benchmark with grace ✔ khuang.build · khuang.validate · kuafai.infer ×27 · kuafai.train ✔ jorfa.score · jorfa.gap · jorfa.facets · jorfa.human · jorfa.errors ✨ Complete · All reports in runs/exp1/

❖

รวมกำเมือง

Complete Index · All Lanna Codenames

Thai	Romanisation	Layer	File	Meaning · Role
กาแล	Kalae	C	kalae_error.py	Guardian gable finial — error analysis & outlier filtering
กัวะ	Kua	B	kua_inference/	Rice cooling tray — inference runner (API + vLLM)
แก้ว	Kaew	C	kaew_aggregator.py	Crystal — Triple-ChrF aggregator (max · avg · diff)
ขันโตก	Khan Tok	B	khantok_registry.py	Serving bowl — 9-model registry
ข้าวตอก	Khao Tok	util	khaotok_seed.py	Sacred popped rice — deterministic seed manager
ข่วง	Khuang	A pkg	lanna_khuang/	Open courtyard — data foundation layer package
คัวไฟ	Kua Fai	B pkg	lanna_kuafai/	Traditional kitchen — adaptation layer package
จ่อฟ้า	Jor Fa	C pkg	lanna_jorfa/	Sky finial — evaluation layer package
ตอก	Tok	A	tok_splitter.py	Bamboo strips — stratified train/dev/test partitioner
ตุง	Tung	A	tung_definer.py	Ritual flag — LannaFacets field definitions
แตะ	Tae	C	tae_slicer.py	Woven bamboo panel — LannaFacets slice analyzer
ไตรภูมิ	Trai Phum	C	traiphum_multiref.py	Three realms — Lanna MultiRef scorer (3 hyp × 3 ref)
น้ำต้น	Nam Ton	A	namton_builder.py	Water carafe — JSONL builder from master sheet
ใบลาน	Bai Lan	C	bailan_form.py	Palm-leaf manuscript — human eval form generator
ปี้	Pee	util	pee_manifest.py	Tag / mark — run manifest (config + hashes + SHA)
ป่อง	Pong	C	pong_gap.py	Gap / interval — Dialect Gap Score G
ป่อหลวง	Po Luang	util	poluang_orchestrate.py	Village headman — pipeline coordinator
ปู่จ๋าน	Pu Jan	util	pujan_io.py	Lay minister — JSONL I/O bridge & hashing
ฟ้อน	Fon	CLI	lanna-fon	Traditional dance — graceful pipeline orchestration
รัก	Rak	A	rak_annotator.py	Lacquer — annotation codebook applicator
สาตุ	Satu	C	satu_robustness.py	Affirmation (amen × 3) — 3P prompt robustness harness
หูก	Huuk	A	huuk_validator.py	Loom — schema validator (Loom-as-gate)
ไหข้าว	Hai Khao	B	haikhao_templater.py	Rice steamer — prompt templater
อั้งโล่	Ang Lo	B	anglo_finetuner.py	Charcoal brazier — LoRA fine-tuner (slow, steady heat)
อูบ	Oob	util	oob_meta/	Ceremonial lacquerware container — shared utilities package
ฮอมปอย	Hom Poi	C	hompoi_agreement.py	Communal merit-making — Krippendorff's α
ฮีต	Heet	util	heet_schema.py	Customary law — Pydantic data schemas (the right way)