Hugging Face's logo

openbmb
/
MiniCPM5-2B

Text Generation
Transformers
Safetensors
English
Chinese
llama
minicpm
minicpm5
long-context
tool-calling
on-device
edge-ai
conversational
text-generation-inference
Model card Files
xet
 Community
18
funmaker commited on
Commit
0e9c66d
·
1 Parent(s): 3497c46

docs: update links of `Just RL II` and formatting in README files for clarity and consistency

Browse files
Files changed (2) hide show
  1. README-cn.md +18 -17
  2. README.md +44 -43
README-cn.md CHANGED
@@ -258,7 +258,7 @@ MiniCPM5-2B 的训练过程是 **[UltraData 分级数据管理体系](https://ar
258
 
259
  ### RL + OPD 带来了什么？
260
 
261
- **RL + OPD** 是 MiniCPM5-2B 后训练中的关键环节。**RL** 阶段，使用了 [JustRL II](https://app.notion.com/p/panhaoxuan/JustRL-II-Scaling-Small-LLMs-to-128K-Reasoning-with-a-Critic-3c77e972297c80adb8b5f4b05d267012#5f3eb56b29f048ebab5f71138f12e36f) 阐述的 critic-based 算法，大幅提升训练稳定性，并在多个领域取得了显著的收益。在下面列出的基准中，RL + OPD 在推理与通用能力上平均提升 **↑ 10.96 分**，Agent 能力平均提升 **↑ 6.96 分**。
262
 
263
  **OPD** 阶段对 16 个 RL 训练所得到的专家模型（含 5 个 agentic 专家模型）实现了能力合并。训练方式上，我们在 response 序列的每个位置分别对学生模型和教师模型 logits 计算全词表的反向 KL 散度作为优势估计值，替代原有的 verification-based advantage；训练数据上，我们的 OPD 直接复用各 RL teacher 训练时 prompt 作为蒸馏数据，无需额外构造语料。
264
 
@@ -359,22 +359,22 @@ MiniCPM5-2B 使用**标准** `LlamaForCausalLM` **架构**，主流推理引擎
359
 
360
  ### 部署
361
 
362
- | 后端 | 模型格式 / 适用场景 | Cookbook | Agent Skill |
363
- | ------------ | ------------------------------------------- | ----------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- |
364
- | Transformers | BF16 / FP16，本地 Python 推理，GPU + CPU | [transformers.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/transformers.md) | [minicpm5-deploy-transformers](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-transformers/SKILL.md) |
365
- | vLLM | BF16 / FP16 OpenAI server | [vllm.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm.md) | [minicpm5-deploy-vllm](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm/SKILL.md) |
366
  | SGLang | BF16 / FP16 OpenAI server，推荐用于 tool calling | [sglang.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/sglang.md) | [minicpm5-deploy-sglang](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-sglang/SKILL.md) |
367
  | llama.cpp | GGUF，CPU/GPU 本地推理 | [llama_cpp.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/llama_cpp.md) | [minicpm5-deploy-llama-cpp](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-llama-cpp/SKILL.md) |
368
- | Ollama | GGUF，本地端侧运行 | [ollama.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/ollama.md) | [minicpm5-deploy-ollama](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-ollama/SKILL.md) |
369
- | LM Studio | GGUF，Mac 桌面应用与 OpenAI server | [lmstudio.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/lmstudio.md) | [minicpm5-deploy-lmstudio](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-lmstudio/SKILL.md) |
370
  | MLX | MLX / 4bit，Apple Silicon 本地推理 | [mlx.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/mlx.md) | [minicpm5-deploy-mlx](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-mlx/SKILL.md) |
371
- | ArcLight | GGUF 本地端侧 / CPU / 桌面 / 服务器 | [arclight.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/arclight.md) | [minicpm5-deploy-arclight](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-arclight/SKILL.md) |
372
- | vLLM Ascend | BF16 / FP16 OpenAI server | [vllm_ascend.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm_ascend.md) | [minicpm5-deploy-vllm-ascend](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm-ascend/SKILL.md) |
373
 
374
  ### 微调
375
 
376
- | 框架 | 适用场景 | Cookbook | Agent Skill |
377
- | ------------- | ------------- | --------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
378
  | TRL + PEFT | LoRA / SFT 微调 | [trl.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/trl.md) | [minicpm5-finetune-trl](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-trl/SKILL.md) |
379
  | LLaMA-Factory | 微调 | [llamafactory.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/llamafactory.md) | [minicpm5-finetune-llamafactory](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-llamafactory/SKILL.md) |
380
  | ms-swift | 微调 | [ms_swift.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/ms_swift.md) | [minicpm5-finetune-ms-swift](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-ms-swift/SKILL.md) |
@@ -399,8 +399,8 @@ FlagOS 社区致力于打造面向多种 AI 芯片的统一、开源的系统软
399
 
400
  基于 FlagOS 极短时间内适配 MiniCPM5-2B 到 9 种不同的 AI 芯片，得益于众智 FlagOS 的多芯片统一 AI 系统软件栈的能力。目前，在 FlagOS 团队构建的面向多架构人工智能芯片的大模型自动迁移、适配与发布平台 FlagRelease 上，已发布 MiniCPM5-2B 的多芯片版本。细节如下：
401
 
402
- | Vendor | ModelScope | Huggingface |
403
- | --------- | ------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- |
404
  | Nvidia | [MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
405
  | Hygon | [MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS) | [MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS) |
406
  | Metax | [MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS) | [MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS) |
@@ -423,8 +423,8 @@ FlagRelease是FlagOS团队构建的一套面向多架构人工智能芯片的大
423
 
424
  ###### FlagRelease 使用速递
425
 
426
- | Vendor | ModelScope | Huggingface |
427
- | --------- | ------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- |
428
  | Nvidia | [MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
429
  | Hygon | [MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS) | [MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS) |
430
  | Metax | [MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS) | [MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS) |
@@ -475,9 +475,10 @@ vllm serve ${model_path} \
475
 
476
  ###### vllm-plugin-FL 使用
477
 
478
- | 厂商 | 从零开始 | 从 FlagRelease 开始 | |
479
- | --- | ---------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------- |
480
  | 英伟达 | [vllm-plugin-FL/MiniCPM5-2B](https://github.com/flagos-ai/vllm-plugin-FL/blob/main/examples/minicpm/README.md) | [MiniCPM5-2B-ModelScope](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
 
481
  </details>
482
 
483
  ## 局限性与免责声明
 
258
 
259
  ### RL + OPD 带来了什么？
260
 
261
+ **RL + OPD** 是 MiniCPM5-2B 后训练中的关键环节。**RL** 阶段，使用了 [JustRL II](https://panhaoxuan.notion.site/justrl-ii-small-llms-to-128k-reasoning-with-a-critic-cn) 阐述的 critic-based 算法，大幅提升训练稳定性，并在多个领域取得了显著的收益。在下面列出的基准中，RL + OPD 在推理与通用能力上平均提升 **↑ 10.96 分**，Agent 能力平均提升 **↑ 6.96 分**。
262
 
263
  **OPD** 阶段对 16 个 RL 训练所得到的专家模型（含 5 个 agentic 专家模型）实现了能力合并。训练方式上，我们在 response 序列的每个位置分别对学生模型和教师模型 logits 计算全词表的反向 KL 散度作为优势估计值，替代原有的 verification-based advantage；训练数据上，我们的 OPD 直接复用各 RL teacher 训练时 prompt 作为蒸馏数据，无需额外构造语料。
264
 
 
359
 
360
  ### 部署
361
 
362
+ | 后端 | 模型格式 / 适用场景 | Cookbook | Agent Skill |
363
+ | ------------ | ------------------------------------------------ | ----------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- |
364
+ | Transformers | BF16 / FP16，本地 Python 推理，GPU + CPU | [transformers.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/transformers.md) | [minicpm5-deploy-transformers](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-transformers/SKILL.md) |
365
+ | vLLM | BF16 / FP16 OpenAI server | [vllm.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm.md) | [minicpm5-deploy-vllm](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm/SKILL.md) |
366
  | SGLang | BF16 / FP16 OpenAI server，推荐用于 tool calling | [sglang.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/sglang.md) | [minicpm5-deploy-sglang](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-sglang/SKILL.md) |
367
  | llama.cpp | GGUF，CPU/GPU 本地推理 | [llama_cpp.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/llama_cpp.md) | [minicpm5-deploy-llama-cpp](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-llama-cpp/SKILL.md) |
368
+ | Ollama | GGUF，本地端侧运行 | [ollama.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/ollama.md) | [minicpm5-deploy-ollama](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-ollama/SKILL.md) |
369
+ | LM Studio | GGUF，Mac 桌面应用与 OpenAI server | [lmstudio.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/lmstudio.md) | [minicpm5-deploy-lmstudio](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-lmstudio/SKILL.md) |
370
  | MLX | MLX / 4bit，Apple Silicon 本地推理 | [mlx.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/mlx.md) | [minicpm5-deploy-mlx](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-mlx/SKILL.md) |
371
+ | ArcLight | GGUF 本地端侧 / CPU / 桌面 / 服务器 | [arclight.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/arclight.md) | [minicpm5-deploy-arclight](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-arclight/SKILL.md) |
372
+ | vLLM Ascend | BF16 / FP16 OpenAI server | [vllm_ascend.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm_ascend.md) | [minicpm5-deploy-vllm-ascend](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm-ascend/SKILL.md) |
373
 
374
  ### 微调
375
 
376
+ | 框架 | 适用场景 | Cookbook | Agent Skill |
377
+ | ------------- | --------------- | --------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
378
  | TRL + PEFT | LoRA / SFT 微调 | [trl.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/trl.md) | [minicpm5-finetune-trl](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-trl/SKILL.md) |
379
  | LLaMA-Factory | 微调 | [llamafactory.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/llamafactory.md) | [minicpm5-finetune-llamafactory](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-llamafactory/SKILL.md) |
380
  | ms-swift | 微调 | [ms_swift.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/ms_swift.md) | [minicpm5-finetune-ms-swift](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-ms-swift/SKILL.md) |
 
399
 
400
  基于 FlagOS 极短时间内适配 MiniCPM5-2B 到 9 种不同的 AI 芯片，得益于众智 FlagOS 的多芯片统一 AI 系统软件栈的能力。目前，在 FlagOS 团队构建的面向多架构人工智能芯片的大模型自动迁移、适配与发布平台 FlagRelease 上，已发布 MiniCPM5-2B 的多芯片版本。细节如下：
401
 
402
+ | Vendor | ModelScope | Huggingface |
403
+ | --------- | --------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
404
  | Nvidia | [MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
405
  | Hygon | [MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS) | [MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS) |
406
  | Metax | [MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS) | [MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS) |
 
423
 
424
  ###### FlagRelease 使用速递
425
 
426
+ | Vendor | ModelScope | Huggingface |
427
+ | --------- | --------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
428
  | Nvidia | [MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
429
  | Hygon | [MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS) | [MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS) |
430
  | Metax | [MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS) | [MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS) |
 
475
 
476
  ###### vllm-plugin-FL 使用
477
 
478
+ | 厂商 | 从零开始 | 从 FlagRelease 开始 | |
479
+ | ------ | -------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------- |
480
  | 英伟达 | [vllm-plugin-FL/MiniCPM5-2B](https://github.com/flagos-ai/vllm-plugin-FL/blob/main/examples/minicpm/README.md) | [MiniCPM5-2B-ModelScope](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
481
+
482
  </details>
483
 
484
  ## 局限性与免责声明
README.md CHANGED
@@ -24,6 +24,7 @@ datasets:
24
  - openbmb/UltraData-SFT-Agent-2609
25
  - openbmb/UltraData-RL-2609
26
  ---
 
27
  <div align="center">
28
  <img src="https://raw.githubusercontent.com/OpenBMB/MiniCPM/main/assets/minicpm_logo.png" width="500em" />
29
  </div>
@@ -268,7 +269,7 @@ During **post-training**, we proceed in three steps: **SFT**, **RL**, and **OPD*
268
 
269
  ### What does RL + OPD bring?
270
 
271
- **RL + OPD** is a key part of MiniCPM5-2B post-training. During the **RL** stage, we adopted the critic-based algorithm described in [JustRL II](https://app.notion.com/p/panhaoxuan/JustRL-II-Scaling-Small-LLMs-to-128K-Reasoning-with-a-Critic-3c77e972297c80adb8b5f4b05d267012#5f3eb56b29f048ebab5f71138f12e36f), substantially improving training stability and achieving significant gains across multiple domains. On the benchmarks listed below, RL + OPD improves reasoning and general capabilities by an average of **↑10.96 points**, and agentic capabilities by **↑6.96 points**.
272
 
273
  **OPD** merges the capabilities of 16 expert models produced by RL training, including 5 agentic expert models. At each response position, we compute the full-vocabulary reverse KL divergence between student and teacher logits as the advantage estimate, replacing the original verification-based advantage. OPD directly reuses the prompts used to train each RL teacher as distillation data, so no additional corpus construction is required.
274
 
@@ -369,26 +370,26 @@ MiniCPM5-2B uses the **standard `LlamaForCausalLM` architecture**, so mainstream
369
 
370
  ### Deployment
371
 
372
- | Backend | Model format / use case | Cookbook | Agent Skill |
373
- | --- | --- | --- | --- |
374
- | Transformers | BF16 / FP16 local Python inference, GPU + CPU | [transformers.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/transformers.md) | [minicpm5-deploy-transformers](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-transformers/SKILL.md) |
375
- | vLLM | BF16 / FP16 OpenAI server | [vllm.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm.md) | [minicpm5-deploy-vllm](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm/SKILL.md) |
376
- | SGLang | BF16 / FP16 OpenAI server, recommended for tool calling | [sglang.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/sglang.md) | [minicpm5-deploy-sglang](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-sglang/SKILL.md) |
377
- | llama.cpp | GGUF local inference, CPU/GPU | [llama_cpp.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/llama_cpp.md) | [minicpm5-deploy-llama-cpp](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-llama-cpp/SKILL.md) |
378
- | Ollama | GGUF local on-device runtime | [ollama.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/ollama.md) | [minicpm5-deploy-ollama](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-ollama/SKILL.md) |
379
- | LM Studio | GGUF Mac desktop app and OpenAI server | [lmstudio.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/lmstudio.md) | [minicpm5-deploy-lmstudio](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-lmstudio/SKILL.md) |
380
- | MLX | MLX / 4bit local inference on Apple Silicon | [mlx.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/mlx.md) | [minicpm5-deploy-mlx](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-mlx/SKILL.md) |
381
- | ArcLight | GGUF local on-device, CPU, Desktop & Server | [arclight.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/arclight.md) | [minicpm5-deploy-arclight](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-arclight/SKILL.md) |
382
- | vLLM Ascend | BF16 / FP16 OpenAI server | [vllm_ascend.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm_ascend.md) | [minicpm5-deploy-vllm-ascend](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm-ascend/SKILL.md) |
383
 
384
  ### Fine-tuning
385
 
386
- | Framework | Use case | Cookbook | Agent Skill |
387
- | --- | --- | --- | --- |
388
- | TRL + PEFT | LoRA / SFT fine-tuning | [trl.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/trl.md) | [minicpm5-finetune-trl](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-trl/SKILL.md) |
389
- | LLaMA-Factory | Fine-tuning | [llamafactory.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/llamafactory.md) | [minicpm5-finetune-llamafactory](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-llamafactory/SKILL.md) |
390
- | ms-swift | Fine-tuning | [ms_swift.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/ms_swift.md) | [minicpm5-finetune-ms-swift](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-ms-swift/SKILL.md) |
391
- | unsloth | Fine-tuning | [unsloth.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/unsloth.md) | [minicpm5-finetune-unsloth](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-unsloth/SKILL.md) |
392
 
393
  ### Other Supported Frameworks
394
 
@@ -409,17 +410,17 @@ Official website express: [https://flagos.io](https://flagos.io/)
409
 
410
  Thanks to FlagOS’s unified multi-chip AI system software stack, MiniCPM5-2B was adapted to 9 different AI chips in an extremely short time. Currently, the multi-chip version of MiniCPM5-2B has been released on FlagRelease, FlagOS’s platform for automatic migration, adaptation, and deployment of large models across multi-architecture AI chips. Details are as follows:
411
 
412
- |Vendor|ModelScope|Huggingface|
413
- |---|---|---|
414
- |Nvidia|[MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS)|[MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS)|
415
- |Hygon|[MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS)|[MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS)|
416
- |Metax|[MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS)|[MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS)|
417
- |Iluvatar|[MiniCPM5-2B-iluvatar-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS)|[MiniCPM5-2B-iluvatar-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS)|
418
- |Zhenwu|[MiniCPM5-2B-zhenwu-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS)|[MiniCPM5-2B-zhenwu-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS)|
419
- |Mthreads|[MiniCPM5-2B-mthreads-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-mthreads-FlagOS)|[MiniCPM5-2B-mthreads-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-mthreads-FlagOS)|
420
- |Kunlunxin|[MiniCPM5-2B-kunlunxin-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS)|[MiniCPM5-2B-kunlunxin-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS)|
421
- |Ascend|[MiniCPM5-2B-ascend-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-ascend-FlagOS)|[MiniCPM5-2B-ascend-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-ascend-FlagOS)|
422
- |ARM-v9|[MiniCPM5-2B-Armv9-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-Armv9-FlagOS)|[MiniCPM5-2B-Armv9-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-Armv9-FlagOS)|
423
 
424
  #### FlagOS Usage
425
 
@@ -433,17 +434,17 @@ FlagRelease is a platform developed by the FlagOS team for automatic migration,
433
 
434
  ###### FlagRelease Quick Start
435
 
436
- |Vendor|ModelScope|Huggingface|
437
- |---|---|---|
438
- |Nvidia|[MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS)|[MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS)|
439
- |Hygon|[MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS)|[MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS)|
440
- |Metax|[MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS)|[MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS)|
441
- |Iluvatar|[MiniCPM5-2B-iluvatar-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS)|[MiniCPM5-2B-iluvatar-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS)|
442
- |Zhenwu|[MiniCPM5-2B-zhenwu-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS)|[MiniCPM5-2B-zhenwu-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS)|
443
- |Mthreads|[MiniCPM5-2B-mthreads-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-mthreads-FlagOS)|[MiniCPM5-2B-mthreads-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-mthreads-FlagOS)|
444
- |Kunlunxin|[MiniCPM5-2B-kunlunxin-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS)|[MiniCPM5-2B-kunlunxin-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS)|
445
- |Ascend|[MiniCPM5-2B-ascend-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-ascend-FlagOS)|[MiniCPM5-2B-ascend-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-ascend-FlagOS)|
446
- |ARM-v9|[MiniCPM5-2B-Armv9-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-Armv9-FlagOS)|[MiniCPM5-2B-Armv9-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-Armv9-FlagOS)|
447
 
448
  ###### From Scratch
449
 
@@ -485,9 +486,9 @@ vllm serve ${model_path} \
485
 
486
  ###### Using vllm-plugin-FL
487
 
488
- |Vendor|From Scratch|From FlagRelease||
489
- |---|---|---|---|
490
- |Nvidia|[vllm-plugin-FL/MiniCPM5-2B](https://github.com/flagos-ai/vllm-plugin-FL/blob/main/examples/minicpm/README.md)|[MiniCPM5-2B-ModelScope](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS)|[MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS)|
491
 
492
  </details>
493
 
 
24
  - openbmb/UltraData-SFT-Agent-2609
25
  - openbmb/UltraData-RL-2609
26
  ---
27
+
28
  <div align="center">
29
  <img src="https://raw.githubusercontent.com/OpenBMB/MiniCPM/main/assets/minicpm_logo.png" width="500em" />
30
  </div>
 
269
 
270
  ### What does RL + OPD bring?
271
 
272
+ **RL + OPD** is a key part of MiniCPM5-2B post-training. During the **RL** stage, we adopted the critic-based algorithm described in [JustRL II](https://panhaoxuan.notion.site/justrl-ii-scaling-small-llms-to-128k-reasoning-with-a-critic), substantially improving training stability and achieving significant gains across multiple domains. On the benchmarks listed below, RL + OPD improves reasoning and general capabilities by an average of **↑10.96 points**, and agentic capabilities by **↑6.96 points**.
273
 
274
  **OPD** merges the capabilities of 16 expert models produced by RL training, including 5 agentic expert models. At each response position, we compute the full-vocabulary reverse KL divergence between student and teacher logits as the advantage estimate, replacing the original verification-based advantage. OPD directly reuses the prompts used to train each RL teacher as distillation data, so no additional corpus construction is required.
275
 
 
370
 
371
  ### Deployment
372
 
373
+ | Backend | Model format / use case | Cookbook | Agent Skill |
374
+ | ------------ | ------------------------------------------------------- | ----------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- |
375
+ | Transformers | BF16 / FP16 local Python inference, GPU + CPU | [transformers.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/transformers.md) | [minicpm5-deploy-transformers](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-transformers/SKILL.md) |
376
+ | vLLM | BF16 / FP16 OpenAI server | [vllm.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm.md) | [minicpm5-deploy-vllm](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm/SKILL.md) |
377
+ | SGLang | BF16 / FP16 OpenAI server, recommended for tool calling | [sglang.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/sglang.md) | [minicpm5-deploy-sglang](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-sglang/SKILL.md) |
378
+ | llama.cpp | GGUF local inference, CPU/GPU | [llama_cpp.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/llama_cpp.md) | [minicpm5-deploy-llama-cpp](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-llama-cpp/SKILL.md) |
379
+ | Ollama | GGUF local on-device runtime | [ollama.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/ollama.md) | [minicpm5-deploy-ollama](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-ollama/SKILL.md) |
380
+ | LM Studio | GGUF Mac desktop app and OpenAI server | [lmstudio.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/lmstudio.md) | [minicpm5-deploy-lmstudio](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-lmstudio/SKILL.md) |
381
+ | MLX | MLX / 4bit local inference on Apple Silicon | [mlx.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/mlx.md) | [minicpm5-deploy-mlx](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-mlx/SKILL.md) |
382
+ | ArcLight | GGUF local on-device, CPU, Desktop & Server | [arclight.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/arclight.md) | [minicpm5-deploy-arclight](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-arclight/SKILL.md) |
383
+ | vLLM Ascend | BF16 / FP16 OpenAI server | [vllm_ascend.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/deployment/vllm_ascend.md) | [minicpm5-deploy-vllm-ascend](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-deploy-vllm-ascend/SKILL.md) |
384
 
385
  ### Fine-tuning
386
 
387
+ | Framework | Use case | Cookbook | Agent Skill |
388
+ | ------------- | ---------------------- | --------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
389
+ | TRL + PEFT | LoRA / SFT fine-tuning | [trl.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/trl.md) | [minicpm5-finetune-trl](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-trl/SKILL.md) |
390
+ | LLaMA-Factory | Fine-tuning | [llamafactory.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/llamafactory.md) | [minicpm5-finetune-llamafactory](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-llamafactory/SKILL.md) |
391
+ | ms-swift | Fine-tuning | [ms_swift.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/ms_swift.md) | [minicpm5-finetune-ms-swift](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-ms-swift/SKILL.md) |
392
+ | unsloth | Fine-tuning | [unsloth.md](https://github.com/OpenBMB/MiniCPM/blob/main/docs/finetune/unsloth.md) | [minicpm5-finetune-unsloth](https://github.com/OpenBMB/MiniCPM/blob/main/skills/minicpm5-finetune-unsloth/SKILL.md) |
393
 
394
  ### Other Supported Frameworks
395
 
 
410
 
411
  Thanks to FlagOS’s unified multi-chip AI system software stack, MiniCPM5-2B was adapted to 9 different AI chips in an extremely short time. Currently, the multi-chip version of MiniCPM5-2B has been released on FlagRelease, FlagOS’s platform for automatic migration, adaptation, and deployment of large models across multi-architecture AI chips. Details are as follows:
412
 
413
+ | Vendor | ModelScope | Huggingface |
414
+ | --------- | --------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
415
+ | Nvidia | [MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
416
+ | Hygon | [MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS) | [MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS) |
417
+ | Metax | [MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS) | [MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS) |
418
+ | Iluvatar | [MiniCPM5-2B-iluvatar-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS) | [MiniCPM5-2B-iluvatar-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS) |
419
+ | Zhenwu | [MiniCPM5-2B-zhenwu-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS) | [MiniCPM5-2B-zhenwu-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS) |
420
+ | Mthreads | [MiniCPM5-2B-mthreads-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-mthreads-FlagOS) | [MiniCPM5-2B-mthreads-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-mthreads-FlagOS) |
421
+ | Kunlunxin | [MiniCPM5-2B-kunlunxin-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS) | [MiniCPM5-2B-kunlunxin-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS) |
422
+ | Ascend | [MiniCPM5-2B-ascend-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-ascend-FlagOS) | [MiniCPM5-2B-ascend-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-ascend-FlagOS) |
423
+ | ARM-v9 | [MiniCPM5-2B-Armv9-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-Armv9-FlagOS) | [MiniCPM5-2B-Armv9-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-Armv9-FlagOS) |
424
 
425
  #### FlagOS Usage
426
 
 
434
 
435
  ###### FlagRelease Quick Start
436
 
437
+ | Vendor | ModelScope | Huggingface |
438
+ | --------- | --------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
439
+ | Nvidia | [MiniCPM5-2B-nvidia-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
440
+ | Hygon | [MiniCPM5-2B-hygon-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-hygon-FlagOS) | [MiniCPM5-2B-hygon-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-hygon-FlagOS) |
441
+ | Metax | [MiniCPM5-2B-metax-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-metax-FlagOS) | [MiniCPM5-2B-metax-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-metax-FlagOS) |
442
+ | Iluvatar | [MiniCPM5-2B-iluvatar-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS) | [MiniCPM5-2B-iluvatar-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-iluvatar-FlagOS) |
443
+ | Zhenwu | [MiniCPM5-2B-zhenwu-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS) | [MiniCPM5-2B-zhenwu-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-zhenwu-FlagOS) |
444
+ | Mthreads | [MiniCPM5-2B-mthreads-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-mthreads-FlagOS) | [MiniCPM5-2B-mthreads-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-mthreads-FlagOS) |
445
+ | Kunlunxin | [MiniCPM5-2B-kunlunxin-FlagOS](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS) | [MiniCPM5-2B-kunlunxin-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-kunlunxin-FlagOS) |
446
+ | Ascend | [MiniCPM5-2B-ascend-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-ascend-FlagOS) | [MiniCPM5-2B-ascend-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-ascend-FlagOS) |
447
+ | ARM-v9 | [MiniCPM5-2B-Armv9-FlagOS](https://modelscope.cn/models/FlagRelease/MiniCPM5-2B-Armv9-FlagOS) | [MiniCPM5-2B-Armv9-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-Armv9-FlagOS) |
448
 
449
  ###### From Scratch
450
 
 
486
 
487
  ###### Using vllm-plugin-FL
488
 
489
+ | Vendor | From Scratch | From FlagRelease | |
490
+ | ------ | -------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------- |
491
+ | Nvidia | [vllm-plugin-FL/MiniCPM5-2B](https://github.com/flagos-ai/vllm-plugin-FL/blob/main/examples/minicpm/README.md) | [MiniCPM5-2B-ModelScope](https://www.modelscope.cn/models/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) | [MiniCPM5-2B-nvidia-FlagOS](https://huggingface.co/FlagRelease/MiniCPM5-2B-nvidia-FlagOS) |
492
 
493
  </details>
494