Instructions to use OpenFormosa/barbet-1b-base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use OpenFormosa/barbet-1b-base with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="OpenFormosa/barbet-1b-base", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("OpenFormosa/barbet-1b-base", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use OpenFormosa/barbet-1b-base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "OpenFormosa/barbet-1b-base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OpenFormosa/barbet-1b-base",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/OpenFormosa/barbet-1b-base

SGLang

How to use OpenFormosa/barbet-1b-base with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "OpenFormosa/barbet-1b-base" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OpenFormosa/barbet-1b-base",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "OpenFormosa/barbet-1b-base" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OpenFormosa/barbet-1b-base",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use OpenFormosa/barbet-1b-base with Docker Model Runner:
```
docker model run hf.co/OpenFormosa/barbet-1b-base
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Barbet 1B Base

OpenFormosa/barbet-1b-base 是 Barbet 1B 基礎模型的 Hugging Face 封裝。模型是 decoder-only hybrid causal LM，交錯使用 global attention、8K sliding-window attention 與 Mamba2 block，詞彙表搭配 openformosa/PangolinTokenizer。

本次發布權重來自實驗保留的 Retry29 iteration 128 checkpoint。它取代先前 iteration 96 的 Hub 權重。

模型架構

項目	設定
Parameters	1,088,124,920
Hidden size	1,536
Logical layers	28
Global-attention layers	7
Sliding-window attention layers	14
Mamba2 layers	7
Attention heads / KV heads	16 / 2
Sliding window	8,192
Vocabulary size	114,944
Weight dtype	BF16

上下文能力與限制

config.json 的 max_position_embeddings=262144 表示 256K 執行路徑，不代表模型已能穩定理解或使用全部 256K tokens。保守的本次發布 claim 是：exact synthetic NIAH 在 32K 穩定通過。

評估	結果	發布判讀
Canonical exact NIAH, 32K	32/32	通過
Canonical exact NIAH, 64K	28/32	未達穩定 gate
Canonical exact NIAH, 128K	25/32	未達穩定 gate
Canonical exact NIAH, 256K	20-23/32	未達穩定 gate
Retry29 easy standard NIAH, native 256K	32/32	機制診斷，不作 capability claim
Retry29 easy opaque NIAH, native 256K	32/32	機制診斷，不作 capability claim
Book-grounded native 256K smoke	5/8	未通過
Clean 320-book evaluation	172/960 (17.92%)	未通過

Exact NIAH 只測量字面定位，不代表 semantic retrieval、證據利用、multi-hop reasoning 或完整長文本理解。研究中的 external/hierarchical verifier 權重也不包含在此 base checkpoint 中。

config_1m_extension.json 提供線性 RoPE scaling factor 4 的研究設定。它只是 experimental 1M runtime configuration，不是原生 1M 訓練，也不是穩定 1M capability claim。

載入方式

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("openformosa/PangolinTokenizer")
model = AutoModelForCausalLM.from_pretrained(
    "OpenFormosa/barbet-1b-base",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Hub 的 config.json 與 config_256k.json 是 256K runtime 設定；config_1m_extension.json 是同一權重的實驗性 1M 外推設定。長序列仍需要相容且經最佳化的執行環境。

CUDA 環境建議安裝相容版本的 mamba_ssm。缺少 mamba_ssm 時，custom model code 會使用較慢的 PyTorch fallback；不同 kernel 與數值精度可能帶來小幅 logits 差異。

發布與轉換

Source checkpoint: retry29_early_pair_retained/r29_earlypair_r0w40v2_srcj150994_i320_j151842_i128
Source iteration: 128
Export format: single BF16 model.safetensors
Tied LM head: lm_head.weight 與 token embedding 共用，未重複儲存
Optimizer、RNG state、raw distributed shards 與 MTP auxiliary tensors未上傳
Conversion details and hashes: conversion_report.json and release_manifest.json

適用範圍

Barbet 1B Base 適合正體中文與多語 causal language modeling、長上下文機制分析及後續微調研究。它不是 instruction-tuned assistant，也未完成面向使用者部署所需的 safety alignment。

已知限制

穩定長度 claim 目前是 32K exact synthetic NIAH，不是 256K 或 1M semantic understanding。
Book-grounded 256K retrieval 與 evidence utilization 尚未通過 release gate。
模型可能生成錯誤、重複、有偏差或不安全內容；高風險用途需要獨立驗證與防護。
License metadata 沿用 other；使用前請確認適用條款與資料授權。

Downloads last month: 1,676

Safetensors

Model size

1B params

Tensor type

BF16

Model tree for OpenFormosa/barbet-1b-base

Finetunes

1 model