Text Generation
Transformers
Safetensors
English
Chinese
qwen3_5
image-text-to-text
qwen
lora
reinforcement-learning
fp8
agentic-coding
scientific-computing
physics
materials-science
genetics
conversational
Instructions to use HoomanBio2021/LabOpen-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use HoomanBio2021/LabOpen-v1 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="HoomanBio2021/LabOpen-v1") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("HoomanBio2021/LabOpen-v1") model = AutoModelForMultimodalLM.from_pretrained("HoomanBio2021/LabOpen-v1", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use HoomanBio2021/LabOpen-v1 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "HoomanBio2021/LabOpen-v1" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HoomanBio2021/LabOpen-v1", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/HoomanBio2021/LabOpen-v1
- SGLang
How to use HoomanBio2021/LabOpen-v1 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 "HoomanBio2021/LabOpen-v1" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HoomanBio2021/LabOpen-v1", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'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 "HoomanBio2021/LabOpen-v1" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HoomanBio2021/LabOpen-v1", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use HoomanBio2021/LabOpen-v1 with Docker Model Runner:
docker model run hf.co/HoomanBio2021/LabOpen-v1
Upload EVALUATION.md with huggingface_hub
Browse files- EVALUATION.md +12 -55
EVALUATION.md
CHANGED
|
@@ -1,60 +1,17 @@
|
|
| 1 |
-
# 能力评估报告 — Qwen3.8-27B-LabOpen v1
|
| 2 |
|
| 3 |
-
|
| 4 |
|
| 5 |
-
|
| 6 |
|
| 7 |
-
|
|
| 8 |
-
|---|---|---|---|---|---|
|
| 9 |
-
|
|
| 10 |
-
|
|
| 11 |
-
|
|
| 12 |
-
|
|
| 13 |
-
|
|
|
|
|
| 14 |
|
| 15 |
-
|
| 16 |
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
| 维度 | 基座 | LabOpen v1 | 判定 |
|
| 20 |
-
|---|---|---|---|
|
| 21 |
-
| 意图理解 | 8/10 | 8/10 | 零回撤 |
|
| 22 |
-
| 智能体编程 | 25/78(32%) | **56/78(72%)** | **+40pp,最大单项净增** |
|
| 23 |
-
| 公式显示 | 5/5 | 5/5 | 零回撤 |
|
| 24 |
-
| 文风(散文) | 2/5 | 0/5* | 见第四节口径说明 |
|
| 25 |
-
|
| 26 |
-
**结论:净增无置换成立**——编程 +40pp、意图/公式零回撤;文风维度存在口径歧义,如实登记。
|
| 27 |
-
|
| 28 |
-
## 三、与外部模型定位
|
| 29 |
-
|
| 30 |
-
LabOpen v1 智能体编程 72% 与 GLM 5.3(73%)几乎并列第一,超过 Qwen3.8-max(69%)与 DeepSeek v4.1 flash(62%)。考虑到 GLM 5.3 为专用大模型而 LabOpen v1 是 27B + r4 LoRA(约 20 个训练窗),训练效率显著。已知差距:
|
| 31 |
-
|
| 32 |
-
- GLM 5.3 首过率 6/10 vs LabOpen v1 5/10——差距集中在首个 agent 用例,一轮修复可解
|
| 33 |
-
- DeepSeek 意图 6/10 为五侧最低——强化侧无此弱点
|
| 34 |
-
|
| 35 |
-
## 四、文风口径说明(诚实登记)
|
| 36 |
-
|
| 37 |
-
LabOpen v1 在本探针 greedy 单抽下 0/5 散文,但收获链评估 T0.3 k=8 采样下 15/20(75%)any-correct 散文。两组数据都真实:
|
| 38 |
-
|
| 39 |
-
- 收获链(T0.3 k=8):分布肩部命中散文 → 15/20
|
| 40 |
-
- 本探针(greedy 单抽):众数走结构化 → 0/5
|
| 41 |
-
- **判读**:文风矫正使散文进入分布但未成为众数——部署口径应推荐 T=0.3 采样(与域计算同理),greedy 下散文不稳定
|
| 42 |
-
- 基座 2/5 是因为其原生散文在众数中占两席
|
| 43 |
-
|
| 44 |
-
## 五、分域能力明细(带标记,T=0.3)
|
| 45 |
-
|
| 46 |
-
| 域 | 指标 | 结果 |
|
| 47 |
-
|---|---|---|
|
| 48 |
-
| 流体物理(CFM) | heldout 池 B,k=8 | 44% 采样命中 |
|
| 49 |
-
| 材料混合律 | heldout 40 项 | 32% 采样命中 |
|
| 50 |
-
| 经典遗传学 | heldout 20 项 | 79% 采样命中(基座自然 47%) |
|
| 51 |
-
| 运维动作查询 | 8 项双温 | 4–5/8 |
|
| 52 |
-
| 散文风格 | 20 项开放题 | 75% any-correct |
|
| 53 |
-
|
| 54 |
-
## 六、诚实边界
|
| 55 |
-
|
| 56 |
-
1. 单种子单检查点,方差带未完全表征
|
| 57 |
-
2. 文风 greedy 不稳定已登记,推荐 T=0.3
|
| 58 |
-
3. 六域能力需带标记寻址,无标记时回退基座行为
|
| 59 |
-
4. 27B 参数量,与 100B+ 模型的通用知识差距如实保留
|
| 60 |
-
5. 合并 FP8 权重相对适配器存在重量化噪声(见 `MERGE_REPORT.md`);适配器为精确数值工件
|
|
|
|
|
|
|
| 1 |
|
| 2 |
+
## 七、域独特性实测(零样本外部对照,2026-10-31)
|
| 3 |
|
| 4 |
+
方法:各域 heldout 原题、剥离标记纯题面、k=8 采样 T=0.3、item any-correct、栈原评分器、输出统一剥 fence。模型配置:GLM 5.3 完整深思;qwen-max 非思考(与智能体 battery 一致);DeepSeek flash 关闭思考。原始数据与完整归因见发布工件 s61-domain-compare(DOMAIN_COMPARE_REPORT.md)。
|
| 5 |
|
| 6 |
+
| 域 | GLM 5.3 | qwen-max | DS flash | LabOpen v1 | 基座 |
|
| 7 |
+
|---|---|---|---|---|---|
|
| 8 |
+
| 流体物理(36) | 36/36 | 36/36 | 35/36 | 25/36† | ≈0 |
|
| 9 |
+
| 材料·严格(40) | 1/40 | 0/40 | 0/40 | 17/40‡ | ≈0 |
|
| 10 |
+
| 材料·公平1%(40) | 20/40 | 15/40 | 22/40 | 18/40‡ | — |
|
| 11 |
+
| 经典遗传(20) | 9/20 | 12/20 | 7/20 | 20/20† | 47% |
|
| 12 |
+
| 运维·严格(8) | 0/8 | 0/8 | 0/8 | 4/8 | 0/8 |
|
| 13 |
+
| 运维·空动作等价(8) | 4/8 | 0/8 | 3/8 | 4/8 | — |
|
| 14 |
|
| 15 |
+
†带标记口径 ‡无标记口径
|
| 16 |
|
| 17 |
+
零分归因:运维严格 0 分含约定伪影(空动作 `observe` 为域私有词表、题面未告知;等价重判后 GLM 4/8、DS 3/8、qwen 0/8)与真实条件绑定失败(前沿以 rollback/restart 直觉替代声明规则、无触发题过度动作);材料严格 0 分为容差伪影(0.01% vs 标准常量,公平口径持平)。诚实表述:运维与遗传为我方独特优势域(条件忠实执行+域词表),材料为语料表精确复现(严格)/物理推理持平(公平),流体物理为本地化能力建设而非独特性。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|