---
license: apache-2.0
base_model: Qwen/Qwen3.8-27B-FP8
tags:
- qwen
- 28B
- Dense
- lora
- reinforcement-learning
- fp8
- agentic-coding
- scientific-computing
- physics
- materials-science
- genetics
language:
- en
- zh
library_name: transformers
pipeline_tag: text-generation
---

# LabOpen v1

Qwen3.8-27B strengthened for verifiable scientific computation — released both as **merged FP8 full weights** (drop-in for the base) and as the original **LoRA adapter**.

LabOpen v1 adds six verifiable capability domains (fluid physics, materials mixture laws, classical genetics, ops action queries, prose style, agentic coding) on top of [Qwen/Qwen3.8-27B-FP8](https://huggingface.co/Qwen/Qwen3.8-27B-FP8), while preserving the base model's native code generation, function calling, intent understanding and formula rendering. The largest measured gain: **agentic coding 32% → 72% (+40pp)** on a 78-test deterministic battery.

- **Base model**: Qwen3.8-27B (FP8 release, Apache 2.0)
- **Training**: 15 reinforcement windows (G69–G84), LoRA r=4, group-relative RL (RLOO-style, k=8 sampling) with mechanically verifiable rewards
- **License**: Apache 2.0 (same as base)
- **Hardware target**: NVIDIA GPUs with compute capability ≥ 8.9 (H100 / L40S / B200 / DGX GB10 class); ~31 GB in FP8

## Positioning & Philosophy

LabOpen v1 is a research model built on the open Qwen3.8-27B foundation through domain fine-tuning and optimization, focused on scientific research fields including programming, physics, life sciences, and materials science.

The project does not aim to replace general-purpose large language models. It seeks to preserve the comprehensive strengths of a general model while further strengthening its adaptation to scientific knowledge, domain-specific problems, and research workloads.

The model is intended for universities, research institutes, corporate R&D centers, and professional laboratories, and supports local or private deployment scaled to the institution's actual computing resources.

## What's in this repo

| Path | Content |
|---|---|
| `model-0000X-of-00007.safetensors` + index | **Merged FP8 full weights** (base + LoRA, blockwise-FP8 re-quantized). Same tensor layout and quantization format as the base — drop-in replacement |
| `adapter/` | **LoRA adapter** (112 MB, PEFT format, 992 tensors). Exact training artifact for inspection and further training — serve the merged weights instead |
| `MERGE_REPORT.md` | Merge provenance: digests, shard SHA-256, re-quantization error stats, output-equivalence probes |
| `EVALUATION.md` | Measured capability matrix and 5-model external comparison |
| `USAGE.md` | Loading, domain tags, temperature guidance, troubleshooting |

Use the **merged weights** for serving (vLLM/SGLang). Use the **adapter** on top of the base when you want the exact training-time numerics or a smaller download.

## Capability matrix (measured, heldout sets)

| Domain | Metric | Score | vs Base |
|---|---|---|---|
| **Agentic coding** | 78 deterministic tests, 3-round fix | **72%** (56/78) | **+40pp** (base: 32%) |
| **Intent understanding** | 10-scene deep-intent probe | **8/10** | preserved |
| **Formula display** | LaTeX integrity, 5 problems | **5/5** | preserved |
| Fluid physics (CFM) | Pool-B heldout, T=0.3 k=8, tagged | 44% draws | new (base ≈ 0) |
| Materials (mixture laws) | heldout 40 items, T=0.3, untagged | 32% draws | new (base ≈ 0) |
| Classical genetics | heldout 20 items, tagged | 79% draws | new (base: 47% natural) |
| Ops (action queries) | 8 items, dual-temp | 4/8 (T0.3) – 5/8 (T0.8) | new |
| Prose style | 20 open-ended, T=0.3 k=8 | 75% any-correct | new |

The matrix reports draw rates (correct samples / total samples) under the deployment protocol. The same heldout sets, scored at item level (any-correct) and compared zero-shot against frontier models, appear in the Domain-capability comparison section below.

### External comparison (same battery, 2026-10)

| Model | Intent | Agentic tests | Rank |
|---|---|---|---|
| GLM 5.3 | 8/10 | 57/78 (73%) | #1 |
| **LabOpen v1 (this)** | **8/10** | **56/78 (72%)** | **#2** |
| Qwen3.8-max | 8/10 | 54/78 (69%) | #3 |
| DeepSeek v4.1 flash | 6/10 | 48/78 (62%) | #4 |
| Qwen3.8-27B (base) | 8/10 | 25/78 (32%) | #5 |

A 27B + r4 LoRA reaching frontier-model agentic-coding territory in 15 training windows is the headline result; see `EVALUATION.md` for gaps that remain (first-pass rate, general knowledge vs 100B+ models).

## Domain-capability comparison with frontier models (zero-shot)

The five external-comparison models above were also run on the **domain heldout sets themselves** (the same questions used for our capability matrix, with tags stripped — plain zero-shot prompts, k=8 samples at T=0.3, item any-correct, scored by the same stack scorers with fence-stripped outputs).

| Domain (items) | GLM 5.3 | Qwen3.8-max¹ | DeepSeek flash | LabOpen v1² | Base |
|---|---|---|---|---|---|
| Fluid physics (36) | 36/36 | 36/36 | 35/36 | 25/36 † | ≈0 |
| Materials — strict (40) | 1/40 | 0/40 | 0/40 | **17/40** ‡ | ≈0 |
| Materials — fair 1% (40) | 20/40 | 15/40 | 22/40 | 18/40 ‡ | — |
| Classical genetics (20) | 9/20 | 12/20 | 7/20 | **20/20** † | 47% |
| Ops — strict (8) | 0/8 | 0/8 | 0/8 | **4/8** (T0.3) | 0/8 |
| Ops — null-action-equivalent (8) | 4/8 | 0/8 | 3/8 | 4/8 | — |

¹ non-thinking mode, consistent with the agentic battery; GLM 5.3 ran with full thinking; DeepSeek flash with thinking disabled (its API option). † our tagged deployment protocol; ‡ our untagged protocol (directly comparable to external).

**Reading the zeros honestly:**

- **Ops 0/8 (strict)** has two components. (a) A convention artifact: the reference null action is the domain vocabulary `"observe"`, which the prompt never states — frontier models answering `"none"` understood the rule but failed the private vocabulary (null-action-equivalent rescoring: GLM 4/8, DeepSeek 3/8, Qwen still 0/8). (b) A genuine failure that remains: frontier models substitute their own ops heuristics for the declared rule — e.g. rolling back on `regression=true` when the declared rule (`escalate_on_critical`, severity below critical) requires no action; across 6 null-trigger items they over-act (rollback/restart dominate their outputs). Our advantage is faithful declared-rule binding plus the domain action vocabulary — not that frontier models "cannot do ops".
- **Materials 0/40 (strict)** is a tolerance artifact: strict scoring (0.01% relative) requires reproducing the corpus property table to 4 significant figures; frontier answers using standard handbook constants (physically correct) fail it. Under a fair 1% tolerance the domain is at parity (38–55% vs our 45%); the strict score measures corpus-table reproduction, our 42% vs their 0–3%.
- **Fluid physics**: frontier models solve these fully zero-shot — our value here is lifting a local 27B from ≈0 to 69%, not uniqueness.

## Usage

### Serving the merged weights (recommended)

```bash
vllm serve HoomanBio2021/LabOpen-v1
# or: docker run ... vllm/vllm-openai --model HoomanBio2021/LabOpen-v1
```

The merged checkpoint uses the same blockwise FP8 (e4m3, 128×128 scales) format as the base — any runtime that loads `Qwen/Qwen3.8-27B-FP8` loads this without changes.

**Hardware support:** requires NVIDIA compute capability ≥ 8.9 (Hopper / Ada / Blackwell — e.g. H100, L40S, B200, DGX GB10; tested on DGX GB10). Ampere and older (A100, RTX 3090) have no FP8 support. Single-card memory must fit ~31 GB of weights (48 GB minimum with reduced context, 80 GB recommended); multi-GPU tensor parallel works. No macOS path — vLLM has no Apple-Silicon backend and no Mac runtime loads this blockwise-FP8 checkpoint format.

**Deployment notes (tested on vLLM 0.27.1):** give domain-computation requests `max_tokens ≥ 4096` — the model reasons before answering, and a 1024 budget is exhausted by the reasoning chain (`finish_reason=length`, empty content). With vLLM's `--reasoning-parser qwen3` the final answer arrives in `content` and reasoning in `reasoning_content`.

### The adapter (research artifact — not a serving path)

The `adapter/` directory holds the exact training artifact (PEFT format, 112 MB) for inspection, provenance and further training. It is **not** a drop-in serving path: plain transformers cannot load the blockwise-FP8 base, and vLLM's LoRA kernels do not cover this architecture's hybrid linear-attention projections (`in_proj_*`, `out_proj`) — 5 of 12 target module types would be silently skipped, corrupting behavior. **Serve the merged weights.**

> **Do not run `merge_and_unload()` on the FP8 base yourself.** The base stores weights as blockwise-FP8 codes with 128×128 scales; a naive PEFT merge adds a BF16 delta directly onto FP8 codes and silently corrupts the weights. Use the merged checkpoint provided here — it was produced by dequantize → add delta → re-quantize with per-block scale recomputation and verified against the adapter (see `MERGE_REPORT.md`).

### Domain tags

The strengthened domains are addressed by a one-line tag at the **start of the prompt**:

| Tag | Domain |
|---|---|
| `[domain: cfm-chain]` | fluid physics chains |
| `[domain: mat-table]` | materials mixture laws |
| `[domain: bio-genetics]` | classical genetics |
| `[domain: ops-action]` | ops action queries |
| `[domain: general-prose]` | open-ended prose style |

Without a tag the model keeps base behavior (code / function calling need no tag).

### Temperature

| Use case | Temperature |
|---|---|
| Domain computation (cfm/mat/bio/ops) | **0.3** (greedy degrades on some domains) |
| Open-ended prose | **0.3** (prose is in the distribution, not always the mode) |
| Code generation | 0.0–0.2 |
| General QA | 0.3–0.7 |

## Technical notes

- **Architecture**: Qwen3.8-27B hybrid — 64 decoder layers: 48 linear-attention + 16 full-attention layers.
- **LoRA**: r=4, α=4 (scaling 1.0), dropout 0, on 496 projection modules across 12 module types (q/k/v/o/gate/up/down_proj, in_proj_a/b/qkv/z, out_proj) = 992 tensors, ~112 MB.
- **Merge math**: per module `W = codes × scale(blockwise)` dequantized to FP32, LoRA delta `(B·A)·scaling` added, re-quantized per 128×128 block (`scale = amax/448`). Mean re-quantization relative error 0.33% — the same order as the base's own FP8 storage noise. Equivalence probes (greedy, 3 prompts × 48 tokens, adapter vs merged): see `MERGE_REPORT.md`.
- **Training substrate**: a local FP8 copy of Qwen3.8-27B (community redistribution). Bit-equivalence with `Qwen/Qwen3.8-27B-FP8` was not verified; the adapter config points to the official release for loading.

## Thinking-behavior governance

The base model can exhaust its thinking budget on unfamiliar domain problems, leaving the reasoning segment unclosed — "thinking exhaustion". This was measured, governed, and re-measured rather than assumed away:

- **Worst recorded case**: before the materials window (G77), 48% of materials generations truncated with unclosed thinking (155/320 during in-window fuel sampling).
- **Mechanism**: the reward contract scores termination mechanically — truncation, format errors, and runtime errors are separate reward categories — and oracle injection teaches termination explicitly. Within that single window, truncation fell from 155/320 to 0/320 and per-item generation time dropped from 327 s to 74 s.
- **Post-governance audit** (four domains, harvest trajectories): unclosed-thinking rate 0% across fluid physics / materials / genetics / ops; degenerate repetition 1/952; median thinking length 180–581 characters (genetics shortest, fluid physics longest).
- **Boundary**: this audit covers the five-domain deployment protocol; open-ended (non-domain) questions have not had a dedicated thinking-exhaustion probe.

## Training methodology (summary)

15 reinforcement windows (G69–G84) on a frozen FP8 serving stack with bit-exact checkpoint accounting: group-relative advantages with k=8 sampling and programmatic answer scoring; oracle injection for zero-mastery domains; verified distillation from champion checkpoints; frozen domain tags; substep advantage for multi-step composition; per-window heldout gates and retention protection. Single seed, single run — no capability claim in this card comes from a set seen during training.

## Honest boundaries

- Single seed, single training run; variance bands not fully characterized
- Prose style is stable at T=0.3 sampling, unstable at greedy decoding
- Domain capabilities require their tags; untagged behavior is base behavior
- 27B parameter count — not comparable to 100B+ models on general knowledge
- Merged FP8 weights add re-quantization noise vs the adapter path (stats in `MERGE_REPORT.md`); the adapter is the numerically exact artifact
- All scores are single evaluation runs on heldout sets

## Acknowledgments

- **Alibaba Qwen team** — for Qwen3.8-27B, the open base model this work builds on, and for the open Apache-2.0 release that makes derivative research like this possible.
- **Z.ai (Zhipu AI)** — GLM 5.3 served as one of the external benchmarks in our evaluation battery and as an engineering assistant during the research and release workflow.
- **Moonshot AI** — Kimi K3 provided engineering assistance in experiment design and testing during the early research process.

We thank these teams for pushing open, measurable frontiers — this release aims to contribute back in the same spirit.

## Citation

```bibtex
@software{labopen_v1,
  title = {LabOpen v1: Six-Domain Reinforcement of Qwen3.8-27B},
  author = {Hooman AI Lab},
  year = {2026},
  url = {https://huggingface.co/HoomanBio2021/LabOpen-v1},
  base = {Qwen/Qwen3.8-27B-FP8},
  license = {Apache-2.0}
}
```

## Contact

If you have any questions, please raise an issue or contact us at [service@hoomanbio.com](mailto:service@hoomanbio.com).

---

# 中文版模型卡

# LabOpen v1

面向可验证科学计算强化的 Qwen3.8-27B——同时以**合并 FP8 全权重**（与基座即插即用）与原始 **LoRA 适配器**两种形态发布。

LabOpen v1 在 [Qwen/Qwen3.8-27B-FP8](https://huggingface.co/Qwen/Qwen3.8-27B-FP8) 之上新增六个可验证能力域（流体物理、材料混合律、经典遗传学、运维动作查询、散文文风、智能体编程），同时保留基座原生的代码生成、函数调用、意图理解与公式显示能力。最大实测增益：78 项确定性智能体编程测试上 **32% → 72%（+40pp）**。

- **基座模型**：Qwen3.8-27B（FP8 版本，Apache 2.0）
- **训练**：15 个强化窗口（G69–G84），LoRA r=4，组相对强化学习（RLOO 式，k=8 采样），机械可验证奖励
- **许可证**：Apache 2.0（与基座相同）
- **硬件要求**：NVIDIA 计算能力 ≥ 8.9 的 GPU（H100 / L40S / B200 / DGX GB10 等级别）；FP8 下约 31 GB

## 项目定位与理念

LabOpen v1 是基于 Qwen3.8-27B 开源基础模型开展领域微调与优化的研究模型，重点面向编程、物理学、生命科学、材料科学等科学研究领域。

项目并不以取代通用大模型为目标，而是希望在保留通用模型综合能力的基础上，进一步增强其对科学专业知识、领域问题及研究任务的适应能力。

模型主要面向高校、科研院所、企业研发中心及专业实验室等使用场景，支持根据实际计算资源条件进行本地化或私有化部署。

## 仓库内容

| 路径 | 内容 |
|---|---|
| `model-0000X-of-00007.safetensors` + index | **合并 FP8 全权重**（基座 + LoRA，分块 FP8 重量化）。张量布局与量化格式和基座完全相同——即插即用替换 |
| `adapter/` | **LoRA 适配器**（112 MB，PEFT 格式，992 张量）。训练精确工件，供检查与继续训练——服务请用合并权重 |
| `MERGE_REPORT.md` | 合并溯源：摘要、分片 SHA-256、重量化误差统计、输出等价探针 |
| `EVALUATION.md` | 实测能力矩阵与五模型外部对比 |
| `USAGE.md` | 加载方式、域标记、温度建议、故障排查 |

部署服务用**合并权重**（vLLM/SGLang）；需要训练时精确数值或更小下载量时用**适配器**加载基座。

## 能力矩阵（heldout 集实测）

| 域 | 指标 | 得分 | 对比基座 |
|---|---|---|---|
| **智能体编程** | 78 项确定性测试，三轮修复 | **72%**（56/78） | **+40pp**（基座：32%） |
| **意图理解** | 10 场景深层意图探针 | **8/10** | 保持 |
| **公式显示** | LaTeX 完整性，5 题 | **5/5** | 保持 |
| 流体物理（CFM） | heldout 池 B，T=0.3 k=8，带标记 | 44% 采样命中 | 新增（基座 ≈ 0） |
| 材料（混合律） | heldout 40 项，T=0.3，无标记 | 32% 采样命中 | 新增（基座 ≈ 0） |
| 经典遗传学 | heldout 20 项，带标记 | 79% 采样命中 | 新增（基座自然 47%） |
| 运维（动作查询） | 8 项，双温 | 4/8（T0.3）–5/8（T0.8） | 新增 |
| 散文文风 | 20 项开放题，T=0.3 k=8 | 75% any-correct | 新增 |

本表报告部署口径下的采样命中率（正确样本/总样本）。同一批 heldout 题以题目级口径（any-correct）与前沿模型零样本对照，见下方「域独特性实测」一节。

### 外部对比（同一测试组，2026-10）

| 模型 | 意图 | 智能体测试 | 排名 |
|---|---|---|---|
| GLM 5.3 | 8/10 | 57/78（73%） | #1 |
| **LabOpen v1（本模型）** | **8/10** | **56/78（72%）** | **#2** |
| Qwen3.8-max | 8/10 | 54/78（69%） | #3 |
| DeepSeek v4.1 flash | 6/10 | 48/78（62%） | #4 |
| Qwen3.8-27B（基座） | 8/10 | 25/78（32%） | #5 |

27B + r4 LoRA 以 15 个训练窗口进入前沿模型的智能体编程区间，这是本发布的头条结果；仍存在的差距（首过率、与 100B+ 模型的通用知识差距）见 `EVALUATION.md`。

## 域独特性实测（与前沿模型零样本对照）

上述外部对比模型同时在**各域 heldout 原题**上复测（同一批题、剥离标记的纯题面、k=8 采样 T=0.3、item any-correct、同一套栈评分器、输出统一剥 fence）。

| 域（题数） | GLM 5.3 | Qwen3.8-max¹ | DeepSeek flash | LabOpen v1 | 基座 |
|---|---|---|---|---|---|
| 流体物理（36） | 36/36 | 36/36 | 35/36 | 25/36 † | ≈0 |
| 材料·严格口径（40） | 1/40 | 0/40 | 0/40 | **17/40** ‡ | ≈0 |
| 材料·公平口径 1%（40） | 20/40 | 15/40 | 22/40 | 18/40 ‡ | — |
| 经典遗传（20） | 9/20 | 12/20 | 7/20 | **20/20** † | 47% |
| 运维·严格口径（8） | 0/8 | 0/8 | 0/8 | **4/8**（T0.3） | 0/8 |
| 运维·空动作等价（8） | 4/8 | 0/8 | 3/8 | 4/8 | — |

¹ 非思考模式，与智能体对比 battery 口径一致；GLM 5.3 保留完整深思；DeepSeek flash 关闭思考（其 API 选项）。† 我方带标记部署口径；‡ 我方无标记口径（与外部直接可比）。

**零分的诚实归因：**

- **运维 0/8（严格）**含两个成分。(a) 约定伪影：参考答案的空动作是域词表 `observe`，题面从未说明——前沿模型答 `none` 说明其理解了规则、只是不掌握私有词表（空动作等价重判：GLM 4/8、DeepSeek 3/8、Qwen 仍 0/8）。(b) 真实失败仍然存在：前沿模型以自身运维直觉替代题面声明的规则——如对 `regression=true` 一律 rollback，而声明规则（escalate_on_critical、severity 未达 critical）应为空动作；6 道无触发题上它们明显过度动作（rollback/restart 占输出主导）。我方优势是声明规则的忠实执行加域动作词表，不是"前沿不会运维"。
- **材料 0/40（严格）**是容差伪影：严格评分（0.01% 相对）要求把语料属性表复现到 4 位有效数字；前沿模型用标准手册常量（物理上正确）被判错。公平 1% 容差下该域持平（38–55% vs 我方 45%）；严格口径衡量的是语料表复现（我方 42% vs 0–3%）。
- **流体物理**：前沿零样本全解——我方价值是把本地 27B 从 ≈0 提到 69%，而非独特性。

## 使用方法

### 部署合并权重（推荐）

```bash
vllm serve HoomanBio2021/LabOpen-v1
# 或：docker run ... vllm/vllm-openai --model HoomanBio2021/LabOpen-v1
```

合并检查点与基座采用完全相同的分块 FP8（e4m3，128×128 块尺度）格式——任何能加载 `Qwen/Qwen3.8-27B-FP8` 的运行时均可直接加载本仓库，无需改动。

**硬件支持**：需要 NVIDIA 计算能力 ≥ 8.9（Hopper / Ada / Blackwell——如 H100、L40S、B200、DGX GB10；已在 DGX GB10 实测）。Ampere 及更早（A100、RTX 3090）无 FP8 支持。单卡显存需容纳约 31 GB 权重（最低 48 GB 并缩短上下文，推荐 80 GB）；支持多卡张量并行。无 macOS 路径——vLLM 无 Apple Silicon 后端，且无 Mac 运行时能加载本分块 FP8 检查点格式。

**部署要点（已在 vLLM 0.27.1 实测）**：域计算请求的 `max_tokens` 应给到 **4096 以上**——模型先思考后作答，1024 预算会被思考链耗尽（`finish_reason=length`、正文为空）。vLLM 配 `--reasoning-parser qwen3` 时，最终答案在 `content`、推理在 `reasoning_content`。

### 适配器（研究工件——非服务路径）

`adapter/` 目录保存训练时的精确工件（PEFT 格式，112 MB），供检查、溯源与继续训练使用。它**不是**即插即用的服务路径：原生 transformers 无法加载分块 FP8 基座；vLLM 的 LoRA 内核不覆盖本架构混合线性注意力投影（`in_proj_*`、`out_proj`）——12 类目标模块中 5 类会被静默跳过、损坏行为。**服务请使用合并权重。**

> **不要自行对 FP8 基座执行 `merge_and_unload()`。** 基座权重以分块 FP8 码 + 128×128 块尺度存储；朴素 PEFT 合并会把 BF16 增量直接加到 FP8 码上，静默损坏权重。请使用本仓库提供的合并检查点——它按「反量化 → 加增量 → 按块重量化」生成，并经与适配器的等价探针验证（见 `MERGE_REPORT.md`）。

### 域标记

强化域通过置于提示首行的单行标记寻址：

| 标记 | 域 |
|---|---|
| `[domain: cfm-chain]` | 流体物理链 |
| `[domain: mat-table]` | 材料混合律 |
| `[domain: bio-genetics]` | 经典遗传学 |
| `[domain: ops-action]` | 运维动作查询 |
| `[domain: general-prose]` | 开放式散文文风 |

不加标记时模型保持基座行为（代码 / 函数调用无需标记）。

### 温度

| 场景 | 温度 |
|---|---|
| 域计算（cfm/mat/bio/ops） | 0.3（部分域 greedy 退化） |
| 开放式散文 | 0.3（散文在分布中但未必是众数） |
| 代码生成 | 0.0–0.2 |
| 通用问答 | 0.3–0.7 |

## 技术说明

- **架构**：Qwen3.8-27B 混合——64 个解码层：48 层线性注意力 + 16 层全注意力。
- **LoRA**：r=4，α=4（缩放 1.0），dropout 0，覆盖 12 类模块（q/k/v/o/gate/up/down_proj、in_proj_a/b/qkv/z、out_proj）共 496 个投影模块 = 992 张量，约 112 MB。
- **合并数学**：逐模块 `W = codes × scale`（分块）反量化至 FP32，加入 LoRA 增量 `(B·A)·scaling`，再按 128×128 块重量化（`scale = amax/448`）。平均重量化相对误差 0.33%——与基座自身 FP8 存储噪声同量级。等价探针（greedy，3 提示 × 48 token，适配器 vs 合并）：见 `MERGE_REPORT.md`。
- **训练基材**：Qwen3.8-27B 的本地 FP8 副本（社区再分发）。未与 `Qwen/Qwen3.8-27B-FP8` 做逐位等价核验；适配器配置指向官方发布以便加载。

## 思考行为治理

基座模型在不熟悉的域问题上可能耗尽思考预算、使推理段未闭合即截断——即"思考枯竭"。本模型对该问题做了测量—治理—复测的完整闭环，而非默认其不存在：

- **最严重记录**：材料域窗口（G77）之前，材料题生成 48% 出现思考未闭合截断（窗内燃料采样期 155/320）。
- **机制**：奖励合同对终止行为做机械评分——截断、格式错误、运行异常是相互独立的奖励类别；oracle 注入显式教授终止行为。该窗口内截断从 155/320 降至 0/320，单题生成时长从 327 秒降至 74 秒。
- **治理后审计**（四域收获轨迹）：流体物理 / 材料 / 遗传学 / 运维四域思考未闭合率均为 0%；退化重复 1/952；思考段长度中位数 180–581 字（遗传学最短、流体物理最长）。
- **边界**：该审计覆盖五域部署口径；开放式（非域）问题尚未做思考枯竭专项探针。

## 训练方法（摘要）

15 个强化窗口（G69–G84），运行于带逐位精确检查点账目的冻结 FP8 服务栈：组相对优势 + k=8 采样 + 程序化答案评分；零起点域的 oracle 注入；冠军检查点的验证蒸馏；冻结域标记；多步组合的子步优势；逐窗 heldout 门与保持保护。单种子单次运行——本卡中的全部能力声明均来自训练未见过的评测集。

## 诚实边界

- 单种子单次训练；方差带未完全表征
- 散文文风在 T=0.3 采样下稳定，greedy 解码下不稳定
- 域能力需要标记；无标记行为即基座行为
- 27B 参数量——通用知识不可与 100B+ 模型相比
- 合并 FP8 权重相对适配器路径存在重量化噪声（统计见 `MERGE_REPORT.md`）；适配器是精确数值工件
- 所有得分均为 heldout 集上的单次评估

## 致谢

- **阿里巴巴 Qwen 团队**——感谢 Qwen3.8-27B 这一开放基座模型，以及 Apache-2.0 开放发布使此类衍生研究成为可能。
- **智谱 AI（Z.ai）**——GLM 5.3 在我们的评估测试组中作为外部基准之一，并在研究与发布流程中担任工程助手。
- **月之暗面（Moonshot AI）**——Kimi K3 在前期研究进程中的实验设计和测试中提供了工程协助。

我们感谢这些团队推动开放、可度量的前沿——本次发布希望以同样的精神回馈社区。

## 引用

```bibtex
@software{labopen_v1,
  title = {LabOpen v1: Six-Domain Reinforcement of Qwen3.8-27B},
  author = {Hooman AI Lab},
  year = {2026},
  url = {https://huggingface.co/HoomanBio2021/LabOpen-v1},
  base = {Qwen/Qwen3.8-27B-FP8},
  license = {Apache-2.0}
}
```

## 联系方式

如有任何问题，请提交问题反馈或联系 [service@hoomanbio.com](mailto:service@hoomanbio.com)。