Skip to page content
SAVRN
Search Plan a project

Open-weight model · Text generation

public-sft-qwen9b-v2

by Alfaplus Laboratory alfaplus/public-sft-qwen9b-v2

public-sft-qwen9b-v2 is an open-weight model for text generation from Alfaplus Laboratory, released under Apache License 2.0. Its published files total 366.3 MB.

本仓库是一个 LoRA 适配器:在 Qwen/Qwen3.5-9B 上,用公开数据集 CohereLabs/ayadataset 的日文子集(languagecode == "jpn")的 2400 条做了一轮指令微调(SFT)。 1.

Parameters—
Context—
Weights346.3 MB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Model Card

By Alfaplus Laboratory, published under apache-2.0, revision 432da6594ab3.

本仓库是一个 LoRA 适配器:在 Qwen/Qwen3.5-9B 上,用公开数据集 CohereLabs/ayadataset 的日文子集(languagecode == "jpn")的 2400 条做了一轮指令微调(SFT)。 1. 几乎全部出自单一标注者:日文 original-annotations 共 6149 行,其中 6144 行出自同一个 userid(9881e959…),另一个标注者只有 5 行(整个日文 original-annotations 只有 2 个标注者)。这些内容里有大量「模板指令 + 一段やさしい日本語体的时事新闻(NHK NEWS WEB EASY 体,2022–2023 年)」,疑似搬自外部文章库而非标注者原创;用 WebSearch 对若干特征句做过核对,既未找到逐字一致的原文、也无法排除(相关旧文多已下线)。 2. 回答普遍很短:训练子集(2400 条)里 assistant 回答 54.2% 在 10 字以内、78.3% 在 50 字以内(多为问答题式的极简短答)。训练后模型很可能偏向极简短答。 3. 未做任何评测,且不得被当成「有日文能力」的模型:本轮只跑了训练,没有任何基准评测,也没有验证集;日文表现完全没有被测过。请不要把本适配器当成一个日文模型——它只是一次在公开日文数据上的小规模 SFT 实验。 不包含:合并后的完整权重、中间 checkpoint、优化器状态、训练语料或其副本。训练数据请从上游获取。 - 数据集:CohereLabs/ayadataset(Singh et al. 2024…

Read Alfaplus Laboratory's full model card

public-sft-qwen9b-v2 · Qwen3.5-9B × Aya Dataset (日文子集) LoRA 适配器(实验产物)

English TL;DR. A LoRA adapter (r=32, alpha=64) for Qwen/Qwen3.5-9B, trained with QLoRA for 3 epochs / 450 optimizer steps on a 2,400-example subset of the Japanese split of CohereLabs/aya_dataset (apache-2.0, revision f9ea04583f02a8f86404ff6c58bf75fe637df8a2, https://huggingface.co/datasets/CohereLabs/aya_dataset). Read the "short-comings" section first: of the 6149 Japanese original-annotations rows, 6144 come from a single annotator (likely sourced from an external article corpus), and 54.2% of the training answers are ≤10 characters. No benchmark evaluation was run — do NOT treat this as a Japanese-capable model. The adapter weights are released under apache-2.0; the dataset's own license is not changed by this release. This is a lab experiment, not a product.

本仓库是一个 LoRA 适配器:在 Qwen/Qwen3.5-9B 上,用公开数据集 CohereLabs/aya_dataset 的日文子集(language_code == "jpn")的 2400 条做了一轮指令微调(SFT)。 只做了训练,没有做任何基准评测。 本卡所有数字都取自这一轮训练的实际产物(训练记录、训练元数据、适配器文件本身)以及对数据的复算;没有估算值。

这份数据/模型的短板(请先读,别软化)

  1. 几乎全部出自单一标注者:日文 original-annotations 共 6149 行,其中 6144 行出自同一个 user_id(9881e959…),另一个标注者只有 5 行(整个日文 original-annotations 只有 2 个标注者)。这些内容里有大量「模板指令 + 一段やさしい日本語体的时事新闻(NHK NEWS WEB EASY 体,2022–2023 年)」,疑似搬自外部文章库而非标注者原创;用 WebSearch 对若干特征句做过核对,既未找到逐字一致的原文、也无法排除(相关旧文多已下线)。
  2. 回答普遍很短:训练子集(2400 条)里 assistant 回答 54.2% 在 10 字以内、78.3% 在 50 字以内(多为问答题式的极简短答)。训练后模型很可能偏向极简短答。
  3. 未做任何评测,且不得被当成「有日文能力」的模型:本轮只跑了训练,没有任何基准评测,也没有验证集;日文表现完全没有被测过。请不要把本适配器当成一个日文模型——它只是一次在公开日文数据上的小规模 SFT 实验。

上面三点在下文「数据」与「已知局限」里再次展开,不是角落备注。

Provenance 一览

项 内容
基座模型 Base model Qwen/Qwen3.5-9B,HF 快照 commit c202236235762e1c871ad0ccb60c8ee5ba337b9a(训练机本地缓存中唯一的快照,refs/main 指向它)。基座自身许可为 apache-2.0
数据集 Dataset CohereLabs/aya_dataset(Aya Dataset v1.0,Cohere Labs 的 Aya Open Science Initiative 众包人工标注),只取日文子集 language_code == "jpn":https://huggingface.co/datasets/CohereLabs/aya_dataset
数据快照 Revision commit f9ea04583f02a8f86404ff6c58bf75fe637df8a2;用其中的 data/train-00000-of-00001.parquet(上游 LFS sha256 51baa85043b569ba117f41516b32d1b0e4e2647fb203a08933f98a556fb1fb06)
数据许可 Data license apache-2.0(上游数据卡 YAML 头 license: 标签逐字如此;正文 "Licensing Information" 一节逐字为 "This dataset can be used for any purpose, whether academic or commercial, under the terms of the Apache 2.0 License.")。原数据集许可不因本发布而改变。
权重许可 Weights license apache-2.0(仅指本仓库的 LoRA 适配器权重)
训练配置 Training QLoRA(基座以 4-bit NF4 加载)· LoRA r32 / α64 / dropout 0.05 / targets auto · seq 1024 · batch 4 × grad_accum 4 · lr 1e-4(cosine)· warmup 0.03 · 3 epoch = 450 步(steps) · seed 42 · bf16 + Liger kernel · 单卡
已知局限 Limitations 单一标注者为主、回答极短、未做任何评测、不代表日文能力;只训 3 epoch / 450 步 / 2400 条;单一 seed;不是产品。详见下文

文件 Files

文件 说明
adapter_model.safetensors LoRA 适配器权重,fp32,496 个张量 / 86,556,672 个参数,346,302,176 字节,sha256 cec0388ca0207247dd00e286e2d7448d657f31b54c3dbff60d234a4f48a06a52(= 第 450 步,也就是最终 checkpoint)
adapter_config.json PEFT 配置(由 peft 0.20.0 写出)
tokenizer.json / tokenizer_config.json / chat_template.jinja 训练时实际使用的 tokenizer 与对话模板,由 transformers 5.15.0 随适配器一起保存(见下文「tokenizer 说明」)

不包含:合并后的完整权重、中间 checkpoint、优化器状态、训练语料或其副本。训练数据请从上游获取。

数据 Data

  • 数据集:CohereLabs/aya_dataset(Singh et al. 2024, arXiv:2402.06619),只取日文子集 language_code == "jpn";来源 URL:https://huggingface.co/datasets/CohereLabs/aya_dataset
  • 快照:revision f9ea04583f02a8f86404ff6c58bf75fe637df8a2,取其中 data/train-00000-of-00001.parquet(上游 LFS sha256 51baa85043b569ba117f41516b32d1b0e4e2647fb203a08933f98a556fb1fb06)。
  • 许可:上游数据卡 YAML 头逐字 license: apache-2.0;正文 "Licensing Information" 逐字为 "This dataset can be used for any purpose, whether academic or commercial, under the terms of the Apache 2.0 License." Apache-2.0 允许再分发与商用,非 NC、非 research-only。原数据集许可不因本发布而改变。
  • 筛选(CPU 复算,口径与训练读取一致):
  • 全语种 202,362 行 → language_code == "jpn":6,259
  • → 只留 annotation_type == "original-annotations":6,149(丢 re-annotations 110 行——那是对开源数据集生成内容的人工改写,其上游各有许可、未逐一核到,故不用)
  • → 去空 0 → 去 inputs 完全重复 3 行(保留首条)→ 可用 6,146
  • → 用 Qwen3.5-9B chat template 实测 token 长度,>1024 丢弃 0 条(全量最长 881 token)
  • → random.Random(42).shuffle 后取前 2,400 条写入训练文件
  • 训练文件:2,400 行,逐行 json.loads 全通过,sha256 3f3500ab8251bc0fc8e414fe26947a7c1c6904f857d64d61d4397a723a5b24f8。
  • 映射:inputs → user,targets → assistant,两轮消息({"messages":[{role,content}...]})。
  • 同源与短答的实测(本次复算):日文 original-annotations 6149 行里 6144 行出自单一 user_id(另一标注者 5 行);训练子集 assistant 回答 54.2% ≤10 字、78.3% ≤50 字。这些内容多为「模板指令 + やさしい日本語时事新闻」或问答题,疑似源自外部文章库;WebSearch 对特征句核对未能逐字证实也未能排除。
  • loss 只算在 assistant 回复 token 上(all_tokens_loss=false)。

训练配置 Training

参数 值
方法 QLoRA:基座经 bitsandbytes 以 4-bit NF4(double quant)加载,计算 dtype bf16;训练 bf16=True
LoRA r=32,lora_alpha=64,lora_dropout=0.05,bias=none,task_type=CAUSAL_LM
target_modules auto:q/k/v/o_proj、gate/up/down_proj、in_proj_qkv/z/a/b、out_proj;exclude_modules 排除 vision/audio/visual/mtp(只覆盖文本解码器)
可训练参数 86,556,672(496 个 fp32 张量)
max_seq_len 1024(训练集中 0 条被截断;最长 881 token)
batch 每卡 4 × grad_accum 4 = 有效 batch 16
optimizer paged_adamw_8bit
lr / schedule 1e-4,cosine;warmup_ratio 0.03
epochs / steps 3 epoch;2400 条 / 有效 batch 16 = 150 步/epoch,共 450 步
seed 42
loss 范围 只算 assistant 回复 token(all_tokens_loss=false)
kernels Liger kernel;attention sdpa;开启 gradient checkpointing
硬件 / 显存峰值 单卡;显存峰值 11.38 GiB
时长 / 吞吐 4,161 秒(约 1.16 小时),219.2 tok/s;每 epoch 304,050 token
软件 transformers 5.15.0、peft 0.20.0
训练日期 2026-10-15

训练集 loss:第 5 步 2.893,第 450 步 0.608。这只是训练集上的 loss,不是验证或评测指标;本轮没有验证集。

怎么用 Usage

import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3.5-9B"
adapter_id = "alfaplus/public-sft-qwen9b-v2"

tok = AutoTokenizer.from_pretrained(adapter_id)
# 用 AutoModelForImageTextToText 加载基座(与训练时一致)。Qwen3.5 的 checkpoint 是 *ForConditionalGeneration,
# 用 AutoModelForCausalLM 加载可能静默漏掉文本解码器权重(只打警告),LoRA 也就挂不对地方。
base = AutoModelForImageTextToText.from_pretrained(
    base_id, revision="c202236235762e1c871ad0ccb60c8ee5ba337b9a",
    dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

msgs = [{"role": "user", "content": "次の文章を一文で要約してください。"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False,
                                 return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
  • 这段示例代码没有在本发布流程中实际运行过(准备发布时不占用 GPU)。加载方式(AutoModelForImageTextToText + PEFT)与训练脚本一致,但请自行验证。
  • 训练样本采用 enable_thinking=False 的格式(空思考块);思考模式下的行为没有评测。
  • 训练时基座是 4-bit NF4 加载的;挂到 bf16 基座上的差异没有测过。

tokenizer 说明(对本仓库实测)

本仓库附带的 tokenizer.json 不是基座 tokenizer.json 的逐字节副本。实测对比本仓库这份与基座那份:

  • 本仓库这份的预切分(pre-tokenizer)正则是 \p{L}+,而基座文件写的是 [\p{L}\p{M}]+(基座含 \p{M} 组合附加符类、本仓库这份不含)。这是 transformers 5.15.0 保存 Qwen3.5 tokenizer 时的已知行为。
  • 行为探针:直接用 tokenizers 库分别读取基座与本仓库的 tokenizer.json,对 10 条多语种/组合字符探针切分,有 3 条不同(都含组合附加符)。
  • 也就是说,训练时实际使用的切分就是本仓库附带的这份 tokenizer。如果推理栈不经过 transformers、直接读取基座的 tokenizer 文件,含组合字符的文本会被切得和训练时不同。

已知局限 Limitations

  • 数据几乎来自单一标注者:日文 original-annotations 6149 行中 6144 行出自同一个 user_id。标注者构成、题材与文体都高度集中,不代表一般日文分布;内容疑似搬自外部文章库(NHK NEWS WEB EASY 体时事新闻、问答题),出处未能逐字证实也未能排除。
  • 回答极短:子集 54.2% ≤10 字、78.3% ≤50 字;训练后模型可能偏向极简短答,不要用它来期待长篇、细致的日文生成。
  • 不代表日文能力,且没有任何评测:只跑了训练,没有做任何基准评测,也没有验证集。上面的 loss 只是训练集 loss,不能说明泛化或与基座的优劣。请勿把本适配器当成日文模型。
  • 训练量小:只训了 3 epoch / 450 步 / 2400 条。
  • 交付形态:交付 LoRA 适配器本体,未合并权重;合并 / 低比特量化(如 GGUF)之后的效果未验证,低比特量化可能削弱适配器的作用。
  • 单一 seed,未做超参搜索。
  • 训练与推理条件:训练时基座 4-bit NF4;其他精度 / 思考模式未测。
  • 只有文本:只训练了文本解码器;挂上适配器后基座的图像 / 视频能力是否受影响未测。
  • 没有任何保证:未做安全评测或红队测试。这是实验室的实验产物,不是产品。

许可 License

  • 适配器权重:apache-2.0。 本仓库 YAML 中的 license: apache-2.0 指适配器权重与本卡。
  • 训练数据:apache-2.0(上游数据卡 YAML 头逐字 license: apache-2.0)。原数据集许可不因本发布而改变。 本仓库不包含 aya 数据或其衍生副本;如需使用数据,请从上游按其许可获取。
  • 基座模型:Qwen/Qwen3.5-9B 按其自身许可(apache-2.0,https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/LICENSE)发布;使用本适配器需另行获取基座并遵守其条款。

引用 / 出处 Citation

@misc{singh2024aya,
      title={Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning},
      author={Shivalika Singh and Freddie Vargus and Daniel Dsouza and others},
      year={2024},
      eprint={2402.06619},
      archivePrefix={arXiv}
}

基座模型:https://huggingface.co/Qwen/Qwen3.5-9B

Identity and Version

Repository
alfaplus/public-sft-qwen9b-v2
Publisher
Alfaplus Laboratory
Task
Text generation
Modality
Text
Library
peft
Parameters
Not stated by the source
Languages
ja
Revision
432da6594ab368e80a777293ff0c89781fcd461c
First published
2026-10-15
Last updated
2026-10-15

Files and Weights

7 files, 366.3 MB in total. The weights are 1 file totalling 346.3 MB in safetensors.

Weights1 file · 346.3 MB
Configuration1 file · 1.3 KB
Tokenizer2 files · 20.0 MB
Documentation1 file · 13.5 KB
Other1 file · 7.8 KB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
adapter_model.safetensorsWeights346.3 MB cec0388ca020
adapter_config.jsonConfiguration1.3 KB —
README.mdDocumentation13.5 KB —
chat_template.jinjaOther7.8 KB —
.gitattributesRepository1.6 KB —
tokenizer.jsonTokenizer20.0 MB 06b9509352d2
tokenizer_config.jsonTokenizer1.1 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
346.3 MB
Download from Alfaplus Laboratory

Released by Alfaplus Laboratory through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published346.3 MB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About public-sft-qwen9b-v2

Can I use public-sft-qwen9b-v2 commercially?

Yes. public-sft-qwen9b-v2 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Text generation

opt-125m

AI at Meta

OPT was first introduced in Open Pre-trained Transformer Language Models and first released in metaseq's repository on May 3rd 2022 by Meta AI. Disclaimer: The team releasing OPT wrote an official model card, which is available in Appendix D of the paper. Content from this model card has been written by the Hugging Face team. To quote the first two paragraphs of the official paper OPT was predominantly pretrained with English text, but a small amount of non-English data is still present within the training corpus via CommonCrawl. The model was pretrained using a causal language modeling (CLM) objective. OPT belongs to the same family of decoder-only models like GPT-3. As such, it was…

Open weights other 2,048 tokens transformers

Model · Text generation

Ornith-1.5-9B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Model · Text generation

Ternary-Bonsai-2-27B-gguf

Prism ML

Full 27B-class reasoning in ternary transformer weights, for llama.cpp (CUDA, Metal, CPU) - \~5.9 GB language model (down from \~54 GB FP16) — full 27B-class reasoning on a standard laptop or a single GPU - 98.2% of FP16 intelligence retained: 84.78 average across 14 thinking-mode benchmarks — far above the conventional IQ2XXS build (72.59) at about 82% of its footprint, and within 0.4 points of UD-Q4KXL at three times the footprint - Retains thinking, reasoning, and agentic behavior deep in the sub-4-bit regime, where conventional low-bit representations collapse: math within half a point of full precision (96.57), coding level with the baseline (89.42), agentic tool calling at 74.92…

Open weights apache-2.0 llama.cpp

Model · Text generation

Ornith-1.5-35B-A3B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Uncensored Qwen3.8-27B, published as GGUF quantizations with the multi token prediction (MTP) head retained and verified. Refusal behaviour has been substantially reduced, not eliminated. See Measured behaviour for the numbers. Capabilities, training data, and architecture are otherwise unchanged. - Refusal directions removed with Heretic, which co minimizes refusal count against KL divergence from the base model. No handwritten refusal removal code, no finetuning, no additional training data. - Abliteration runs at bf16 (no 4 bit quantization). the resulting LoRA is merged into the bf16 base, so the published weights are not a quantized round trip. - mtp. tensors are copied verbatim from…

Open weights apache-2.0 llama.cpp

Model · Text generation

Ornith-1.0-9B-GGUF

Ornith

Aloha! Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding. This model card documents Ornith-1.0-9B, the most lightweight member of the Ornith family, designed for efficient single-GPU deployment. Ornith-1.0-9B is a dense ~9B model (≈19 GB in bf16), so it serves comfortably on a single 80GB GPU. The recipes below stand up an OpenAI-compatible server; add --tensor-parallel-size / --tp if you want to shard across more GPUs. For a quick local test (or to script offline generation), load the model directly with Transformers. Make sure you have a recent release installed — see the Transformers installation guide; Ornith-1.0-9B requires…

Open weights mit transformers