Couplet

Name: Couplet
Availability: InStock

Instruction-tuned chat model (~268M) — Couplet family

Stage

License

Base

A ~268M hybrid GDN-2 + GQA model, supervised fine-tuned for chat

What this is

Couplet is the supervised fine-tuned (SFT) chat checkpoint for the Couplet family.

Base (pretrain): kerzgrr/Couplet-base
Larger successor to kerzgrr/Monostich-2
Architecture: hybrid Gated DeltaNet-2 + gated GQA

Training

Pretrain → SFT

Stage	Details
Base	5.47B-token FineWeb-Edu pretrain → `Couplet-base`
SFT mix	HuggingFaceTB/smoltalk + NousResearch/Hermes-3-Dataset
Epochs	1 full epoch
Assistant targets	977,656,016
Wall time	12.47 hours
Final step	optimizer step 10,619
Weights	EMA (Hub `model.safetensors` is EMA @ bfloat16)
Seq length	8,192 (packed SFT)
Peak LR	1 × 10⁻⁴ AdamW, cosine → 10% min
Final val loss (EMA)	1.2862 (ppl 3.62)

Chat template (ChatML)

<|begin_of_text|><|im_start|>system
{system}<|im_end|>
<|im_start|>user
{user}<|im_end|>
<|im_start|>assistant
{assistant}<|im_end|>

Generation prompt ends at <|im_start|>assistant\n.

Model Architecture

Same TinyGDN hybrid as the base (267,857,662 parameters):


Layers	28 (GDN-2 ×3 + GQA every 4th)
Hidden	768
MLP	SwiGLU 2,048
Attention	6 Q / 2 KV, head dim 128, partial RoPE
Linear	Gated DeltaNet-2, 6 heads × 128
Vocab	49,152 BPE

IFEval

Official 541-prompt Google IFEval scorer, zero-shot ChatML, greedy decoding (temperature=0, max_new_tokens=1280):

Metric	Score
Prompt-level strict	16.3% (88 / 541)
Instruction-level strict	28.4% (237 / 834)
Prompt-level loose	18.5% (100 / 541)
Instruction-level loose	31.4% (262 / 834)

Install & run

bash

pip install torch safetensors tokenizers huggingface_hub
hf download kerzgrr/Couplet inference.py --local-dir .
python inference.py --prompt "What is the capital of France?"

inference.py auto-downloads weights/tokenizer/tiny_gdn/ and auto-installs pinned flash-linear-attention (Windows applies Hub patches). Git is required on PATH.

Interactive chat:

python inference.py

Flag	Default	Description
`--prompt`	—	One-shot user message
`--system`	—	Optional system prompt
`--temperature`	`0.7`	Sampling temperature
`--top-p`	`0.9`	Nucleus sampling
`--top-k`	`50`	Top-k
`--max-new-tokens`	`256`	Max generation length
`--device`	`cuda` if available	`cuda` / `cpu`

Limitations

Scale: at ~268M parameters this is a research / edge model, not a frontier system
Instruction following: IFEval shows useful but limited constraint following; verify critical outputs
Dependency: requires flash-linear-attention; not GGUF / llama.cpp compatible today

Model family

Model	Stage	Hub
Monostich	SFT (~100M LLaMA)	`kerzgrr/Monostich`
Monostich-2-base	Pretrain (~150M hybrid)	`kerzgrr/Monostich-2-base`
Monostich-2	SFT (~150M hybrid)	`kerzgrr/Monostich-2`
Couplet-base	Pretrain (~268M hybrid)	`kerzgrr/Couplet-base`
Couplet	SFT (~268M hybrid)	this repo

Citation

bibtex

@misc{couplet2026,
  title={Couplet: A 268M Hybrid GDN-2 + GQA Chat Model},
  author={kerzgrr},
  year={2026},
  url={https://huggingface.co/kerzgrr/Couplet}
}

Acknowledgments

flash-linear-attention (Gated DeltaNet-2)
HuggingFaceTB/smoltalk
NousResearch/Hermes-3-Dataset
Base: kerzgrr/Couplet-base

A couplet is a pair of lines — compact, connected, and complete.