README.md

---
base_model: Qwen/Qwen3-4B-Base
library_name: peft
tags:
- peft
- lora
- qwen3
- math
- evalscope
---
# MT11-HIGH-L45 — v1.0.0

## 中文说明

这是 `HIGH_HEAVY_L45` 的**未合并动态 LoRA** 发布件：LoRA rank `r=64`、`alpha=128`，base 与 adapter 分离，未 merge、未 bake。请勿把本仓库当作完整基础模型。

### 基础模型与固定身份

- 基础模型：`Qwen/Qwen3-4B-Base`
- 训练时精确本地路径：`/workspace/MATH-NOTHINK-AIME-P1-v1.0/model/mn8_eos_r1_2/Qwen3-4B-Base-906bfd4`
- 固定 revision：`906bfd4b4dc7f14ee4320094d8b41684abff8539`
- base composite SHA256：`b7dc3d5cef56c5ed8e03e9c54bde781b560897151e6b6fadc55ed40a22feaaa2`
- base lock SHA256：`0144c9767921f51f37167784380029c01fc83caf5c5b54e803671bfc3c432bdf`
- tokenizer：使用上述精确 base revision 自带 tokenizer；本仓库不重复上传 tokenizer 文件。
- adapter SHA256：`628a5b4c40584c524479b80773166f49b5c49d7223b34d4cf254064d6b65bd29`

### 评测合同与结果

MT11 标准合同为 **Think + EvalScope reviews + dynamic LoRA + 32K**：AIME24+AIME25 共 60 题，temperature 0.6、top_p 0.95、top_k 20，输出上限为 `32768 - exact_prompt_tokens - 64`，统计以题目为 cluster，不把 240 个采样视为 240 道独立题。

- seeds 42–45：`34/240`
- fresh seeds 46–49：`34/240`
- combined seeds 42–49：`68/480`（14.167%）
- fresh health：cap 35.000%，decoded complete box 62.917%，natural stop 65.000%
- combined health：cap 36.458%，decoded complete box 62.708%，natural stop 63.542%

### 来源完整性披露

状态为 **`ASSET_FROZEN_PROVENANCE_SCHEMA_MISSING`**。缺失的是训练当时 SHA256 为 `1819b4a0abf92ac070cf8bb84a47b4a7b8e72a537df2bee81687ae8f1a31a527` 的 schema 字节副本；现环境 schema 只能作为环境快照，不能冒充训练时副本。adapter 字节、adapter config/COMMIT 绑定及两组历史评测已独立核验，但**不保证精确重训复现**。

### 动态 LoRA 使用示例

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-4B-Base"
base_revision = "906bfd4b4dc7f14ee4320094d8b41684abff8539"
adapter_id = "modrill/MT11-HIGH-L45"
revision = "v1.0.0"

tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)
base = AutoModelForCausalLM.from_pretrained(base_id, revision=base_revision, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id, revision=revision)  # dynamic, unmerged
```

### 数据与许可证限制

本发布不包含训练数据、原始 responses、trainer state 或 optimizer。现有冻结证据未授权为该 adapter 编造或新增独立许可证，因此本仓库**不声明新的 adapter license**。使用者必须自行核对并遵守基础模型、训练数据来源及适用数据/竞赛内容条款；评测结果不构成许可授权。

## English

This is the **unmerged dynamic LoRA** release of `HIGH_HEAVY_L45` (`r=64`, `alpha=128`). The base and adapter stay separate; nothing is merged or baked.

The exact base is `Qwen/Qwen3-4B-Base` at revision `906bfd4b4dc7f14ee4320094d8b41684abff8539` (training path and hashes are listed above). Use that revision's tokenizer. The adapter SHA256 is `628a5b4c40584c524479b80773166f49b5c49d7223b34d4cf254064d6b65bd29`.

Under the MT11 Think / EvalScope / dynamic 32K contract, seeds 42–45 scored 34/240, fresh seeds 46–49 scored 34/240, and the combined block scored 68/480. Fresh health: 35.000% cap, 62.917% decoded-complete-box, 65.000% natural stop. Combined health: 36.458%, 62.708%, and 63.542%, respectively.

Provenance status is **`ASSET_FROZEN_PROVENANCE_SCHEMA_MISSING`**: the training-time schema byte copy is missing. The available current-environment schema is not represented as the historical training schema. Adapter bytes and evaluation evidence were verified, but exact retraining reproducibility is **not guaranteed**.

No training data, raw responses, trainer state, optimizer state, or secrets are included. Existing project evidence does not establish a new standalone license for this adapter, so none is invented here. Users must review and comply with the base-model and applicable data/source terms.