songgot-l

Songgot (송곳)

A Korean-first tiny agentic model for tool calling on the device, trained from scratch by Hanish Keloth (Palette). Apache 2.0.

• Paper: https://hanishkeloth.github.io/songgot
• Code, data generators, scorer: https://github.com/hanishkeloth/songgot
• Demo: https://huggingface.co/spaces/Hanish/songgot

Try it on the device: https://hanishkeloth.github.io/songgot/app/ (runs in the browser, works offline after the first load).

Numbers

Kakao FunctionChat-Bench SingleCall (500 Korean items, 5 tool conditions), exact match on function name and arguments, scorer in the repo. Comparators run with identical tools and queries in their own documented formats.

Tokens per Hangul syllable on the same 100 queries: Songgot 0.90, Gemma 3 0.98, Qwen3 1.15, Needle 2 3.47.

Status (2026-09-13 09:33)

Weights in this repo are Songgot-L, 12B tokens with instruction bucket, 2 epochs, v8 set: 24 layers, hidden 1024, about 303M parameters, trained from scratch on 8xH100 (Modal) on 12B tokens of Korean Wikipedia and fineweb-edu with 5 percent tool-calling rows in the mix, post-trained on the v8 set, post-trained on the v2 tool-calling set. Call accuracy on FunctionChat-Bench SingleCall 29.4 percent (name only 75.0). GGUF exports (f16, Q8_0, Q4_K_M) are in this repo.

Format

<|system|>
[{"name": "set_alarm", "description": "알람을 설정합니다.", "parameters": {...}}]
<|user|>
내일 아침 7시에 알람 맞춰줘
<|call|>
{"name":"set_alarm","arguments":{"time":"07:00"}}<|end|>

Tokenizer: SentencePiece BPE, 32k, byte fallback (tokenizer.model). Use sentencepiece directly; the special tokens live inside the vocabulary.

Data and provenance

fineweb-edu sample-10BT (ODC-By), Korean Wikipedia 20231101.ko (CC BY-SA 3.0; this model card carries the attribution and share-alike notice for that text), glaive-function-calling-v2 (Apache 2.0), template-generated Korean tool calls (Apache 2.0, in the repo). No closed-model outputs. FunctionChat-Bench was never used for training.

Limits

Single-call tool selection and argument extraction only. No multi-turn, no tool results, no free chat. Small models are finicky with rare tools and paraphrased values; validate every call in application code.