初始化项目，由ModelHub XC社区提供模型

Model: HansBug/Qwen3-30B-A3B-OpenClaw-RL-iter335 Source: Original Platform
2026-09-26 19:24:01 +08:00
commit f1e869b047
26 changed files with 171148 additions and 0 deletions
--- a/.gitattributes
+++ b/.gitattributes
@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text
+training_curves.png filter=lfs diff=lfs merge=lfs -text
--- a/202
+++ b/202
@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2024 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
--- a/README.md
+++ b/README.md
@@ -0,0 +1,314 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-30B-A3B
+tags:
+- qwen3
+- qwen3-moe
+- terminal-agent
+- reinforcement-learning
+- grpo
+- camel-ai
+- openclaw-rl
+- agent
+- terminal-bench
+- moe
+library_name: transformers
+language:
+- en
+- zh
+pipeline_tag: text-generation
+---
+
+# Qwen3-30B-A3B-OpenClaw-RL-iter335
+
+把 [`Qwen/Qwen3-30B-A3B`](https://huggingface.co/Qwen/Qwen3-30B-A3B) (vanilla, 30B total / 3B active MoE) 用 [`HansBug/OpenClaw-RL`](https://github.com/HansBug/OpenClaw-RL) 的训练框架做 **GRPO outcome-only RL**（dense pass-rate reward，无 PRM、无 process reward）跑出来的 **best 中间 checkpoint**，对应 [wandb run `b0il0mq4`](https://wandb.ai/hansbug/openclaw-terminal-rl/runs/b0il0mq4) 的 rollout 335 (megatron iteration 335)。
+
+在 [`terminal-bench` v0.1.x](https://github.com/laude-institute/terminal-bench/tree/v0.1.x) 66 个 OOD task 上把 base Qwen3-30B-A3B 的 pass@1 从 0.0606 推到 **0.0808** (+33%)、pass@3 从 0.1212 推到 **0.1515** (+25%)，是这次 30B fresh run 训练中**唯一明确"超越自身 base"的 ckpt**。
+
+> **完全 drop-in 兼容**：config / generation_config / tokenizer 全部跟上游 `Qwen/Qwen3-30B-A3B` 字节级一致；任何能跑 Qwen3-30B-A3B 的工具链（HF transformers、vLLM、sglang、ollama、llama.cpp 转 GGUF 等）**一行都不用改**。
+
+完整的训练 + eval 详细分析见 [`HansBug/OpenClaw-RL` issue #14](https://github.com/HansBug/OpenClaw-RL/issues/14)。
+
+---
+
+## 目录
+
+1. [快速开始](#快速开始)
+2. [模型从哪来](#模型从哪来)
+3. [训练设置](#训练设置)
+4. [训练曲线](#训练曲线)
+5. [评测结果](#评测结果)
+6. [为什么是 iter_335 而不是最新 ckpt](#为什么是-iter_335-而不是最新-ckpt)
+7. [推理时的 on-the-wire 协议](#推理时的-on-the-wire-协议)
+8. [已知限制](#已知限制)
+9. [复现 / 推理工具](#复现--推理工具)
+10. [引用 / 致谢](#引用--致谢)
+
+---
+
+## 快速开始
+
+### 用 HF `transformers`
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tok = AutoTokenizer.from_pretrained("HansBug/Qwen3-30B-A3B-OpenClaw-RL-iter335")
+model = AutoModelForCausalLM.from_pretrained(
+    "HansBug/Qwen3-30B-A3B-OpenClaw-RL-iter335",
+    torch_dtype="bfloat16",
+    device_map="auto",
+)
+
+messages = [
+    {"role": "user", "content": "Write a one-line bash command to recursively count *.py files under /usr/lib."}
+]
+text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+inputs = tok(text, return_tensors="pt").to(model.device)
+out = model.generate(**inputs, max_new_tokens=256, temperature=0.2)
+print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
+```
+
+### 用 sglang（推荐，配 tool-call 支持 + MoE 加速）
+
+```bash
+python -m sglang.launch_server \
+    --model-path HansBug/Qwen3-30B-A3B-OpenClaw-RL-iter335 \
+    --served-model-name qwen3-30b-rl-iter335 \
+    --tp 4 --port 30000 \
+    --mem-fraction-static 0.85 \
+    --tool-call-parser qwen \
+    --disable-custom-all-reduce
+```
+
+跟 base Qwen3-30B-A3B 一样的启动参数。TP=4 时 30B MoE 单 rank ~15 GB params + ~120 GB kv cache，放 4× 80 GB GPU 或 4× 143 GB GPU 都没问题。`--disable-custom-all-reduce` 是 [issue #8](https://github.com/HansBug/OpenClaw-RL/issues/8) 已知 Hopper bug 的兜底 flag，保留即可。`--tool-call-parser qwen`（或 `qwen25`）把模型的 `<tool_call>{...}</tool_call>` 还原成 OpenAI 结构化 `tool_calls`，这正是训练时 rollouts 用的 parser。
+
+### 用 vLLM
+
+```bash
+vllm serve HansBug/Qwen3-30B-A3B-OpenClaw-RL-iter335 \
+    --tensor-parallel-size 4 \
+    --enable-auto-tool-choice --tool-call-parser hermes
+```
+
+### 用 oc-repl 直接交互（端到端 demo）
+
+[`HansBug/oc-repl`](https://github.com/HansBug/oc-repl) 是一个针对这个 family 量身写的 Codex 风格 REPL：
+
+```bash
+pip install git+https://github.com/HansBug/oc-repl
+
+# 起一个 sandbox 容器（任何 ubuntu/python image 都行）
+docker run -d --name oc-sandbox -w /app ubuntu:24.04 sleep infinity
+
+# 启动交互式 REPL
+oc-repl --sandbox docker:oc-sandbox \
+        --api-base http://127.0.0.1:30000/v1 \
+        --model qwen3-30b-rl-iter335
+```
+
+oc-repl 默认 `--protocol camel-terminal-toolkit` 对齐训练分布，详见仓库 README。
+
+---
+
+## 模型从哪来
+
+`Qwen3-30B-A3B-OpenClaw-RL-iter335` = `Qwen3-30B-A3B` (vanilla, **不是** Coder-30B-A3B-Instruct，也**不是** 30B-A3B-Instruct-2507) + **335 个 RL rollout** 的 GRPO outcome-only 训练，rollout 在 [SETA terminal env pool](https://github.com/camel-ai/seta) 提供的 **1376 个 terminal task** 上。每个 task 是一个真实的 Linux shell 任务（修脚本、解 base64、起 server、跑 pytest…），在隔离 docker 容器里跑，agent 通过 [`camel.toolkits.TerminalToolkit`](https://docs.camel-ai.org/key_modules/terminaltoolkit) 暴露的 4 个工具操作 shell。
+
+| | 值 |
+|---|---|
+| 基座 | [`Qwen/Qwen3-30B-A3B`](https://huggingface.co/Qwen/Qwen3-30B-A3B)（**vanilla** MoE，30B total / 3B active，128 expert / topk=8 / 48 layer / hidden 2048）|
+| 训练框架 | [`HansBug/OpenClaw-RL`](https://github.com/HansBug/OpenClaw-RL) (Megatron + slime + sglang rollout) |
+| Agent harness | [camel-ai `ChatAgent` + `TerminalToolkit`](https://docs.camel-ai.org/key_modules/terminaltoolkit) |
+| 数据 | [seta_env](https://github.com/camel-ai/seta) 1376 task pool（含 [terminal-bench](https://github.com/laude-institute/terminal-bench) v0.1.x 86 task 子集） |
+| 算法 | GRPO outcome-only，dense pass-rate reward |
+| 训练 rollout 数 | 335 rollout（约 4× n_samples × prompt 数） |
+| 训练硬件 | 8× NVIDIA H200，actor 4×TP + 4×EP + ETP=1，rollout 4×TP |
+| 训练时长 | 约 4 天到 iter_335 完成（完整 fresh run 直到 iter_367 共 ~5.5 天 / 367 rollout）|
+| Wandb | [`hansbug/openclaw-terminal-rl/runs/b0il0mq4`](https://wandb.ai/hansbug/openclaw-terminal-rl/runs/b0il0mq4) |
+| Group | `qwen3-30b-a3b-vanilla-fresh-run1` |
+
+**为什么是 vanilla 而不是 Coder-30B-A3B-Instruct？** 初次用 Coder-30B-A3B-Instruct 跑了 1114 个 rollout 全 fail —— 它的 `chat_template.jinja` 用 `tool_call.arguments|items` 要求 `arguments` 是 dict，但 slime/sglang 走 OpenAI 协议把 `arguments` 序列化成 JSON string → jinja2 `TypeError`。vanilla `Qwen3-30B-A3B` 的 chat_template 与 8B 字节级相同，整条 slime + terminus-2 + harbor 链路 0 修改通过。详见 [`HansBug/OpenClaw-RL` issue #14 §1.1](https://github.com/HansBug/OpenClaw-RL/issues/14)。
+
+---
+
+## 训练设置
+
+完整启动命令见 [`run_qwen3_30b_a3b_experiment.sh`](https://github.com/HansBug/OpenClaw-RL/blob/main/terminal-rl/run_qwen3_30b_a3b_experiment.sh)。关键超参（与 8B run-3 算法侧完全相同，仅 MoE 并行差异）：
+
+```bash
+# GRPO 算法 (同 8B run-3, issue #4)
+--advantage-estimator grpo
+--use-kl-loss --kl-loss-coef 0.01 --kl-loss-type k3
+--dynamic_history
+
+# Optimizer (low LR + Adam beta2=0.98)
+--optimizer adam --lr 1e-6 --lr-decay-style constant
+--weight-decay 0.1 --adam-beta1 0.9 --adam-beta2 0.98
+
+# Rollout (同 8B)
+--rollout-batch-size 16
+--n-samples-per-prompt 8           # GRPO group size = 8
+--num-steps-per-rollout 2
+--rollout-temperature 1
+--rollout-max-response-len 8192
+--rollout-max-context-len 16384
+
+# Megatron — MoE 并行（与 8B 不同）
+--tensor-model-parallel-size 4
+--expert-model-parallel-size 4     # 128 expert 分到 4 张 actor 卡，每卡 32 expert
+--expert-tensor-parallel-size 1
+--sequence-parallel
+--recompute-granularity full
+
+# Save / freq
+--save-interval 16                  # save 每 16 rollout（≈ 5.3h/save）
+--num-rollout 2000                  # 训练目标（实际跑到 367 停）
+
+# Rollout agent + tool-call parser
+--custom-config-path configs/rollout_qwen3.yaml   # tool_call_parser: qwen25
+                                                  # max_iteration: 10
+```
+
+Reward signal：**dense pass-rate, outcome-only** —— 每个 rollout 跑完后由任务自带的 `run-tests.sh` 在 sandbox 里跑 pytest，`reward = passed / total_tests ∈ [0, 1]`，再线性变换到 `score = 2·reward − 1 ∈ [-1, +1]` 喂给 GRPO；**没有 PRM / process reward**，整个 episode 共用一个标量。
+
+`max_iteration: 10`：训练时 agent loop 上限 10 轮。
+
+---
+
+## 训练曲线
+
+来自 [wandb run b0il0mq4](https://wandb.ai/hansbug/openclaw-terminal-rl/runs/b0il0mq4)：
+
+![training curves](training_curves.png)
+
+六个子图说明：
+
+| 子图 | 解读 |
+|---|---|
+| **raw_reward** | rollout 阶段 reward 均值。-0.07 (rollout 0-49) → +0.21 (rollout 320-335) 单调上升。**穿越 0 出现在 rollout 50 左右**，即模型从"过的 test < 一半"转向"过的 test > 一半"。iter_335 这个 ckpt 对应的训练窗口 raw_reward 均值 **+0.21**，比 base 提升 ~0.28。|
+| **response_length** | 每轮 assistant 输出的平均 token 数。从 217 涨到 413，**单调上升、没有 plateau**——是 30B 远未到达容量上限的关键信号（vs 8B run-3 后期横盘 300-400）。|
+| **grad_norm** | 梯度幅度（log 尺度）。median 0.30-0.50 健康，期间 5 次单步 spike（最大 step 38 grad=1.36e8 / kl=5.6e5），全部被 megatron `--clip-grad 1.0` 兜底，下一步即恢复。|
+| **kl_loss** | K3 KL estimator。median 0.05-0.15 健康，远低于 0.5 阈值。spike 与 grad_norm 同步出现，clip-grad 同样吸收。|
+| **entropy_loss** | 0.13-0.17 持续，无 collapse。|
+| **pg_clipfrac** | PPO clip 比例 ~0.6%（合理）。|
+
+详细的训练曲线、异常事件分析、vs 8B run-3 叠加图见 [`HansBug/OpenClaw-RL` issue #14](https://github.com/HansBug/OpenClaw-RL/issues/14)。
+
+---
+
+## 评测结果
+
+### 在 terminal-bench v0.1.x（66-task harbor 子集）上的 pass@1 / pass@3
+
+iter_335 这个 ckpt 在 OOD eval（agent harness 切到 `harbor run --agent terminus-2`，与 [issue #8](https://github.com/HansBug/OpenClaw-RL/issues/8) 8B run-3 同协议）上的成绩：
+
+![eval pass rate](eval_pass_rate.png)
+
+| Model | pass@1 | pass@3 | unique solved (/66) | solved trials (/198) |
+|---|---:|---:|---:|---:|
+| `qwen3-30b-a3b-base` (vanilla) | 0.0606 | 0.1212 | 8 | 12 |
+| **`qwen3-30b-a3b-OpenClaw-RL-iter335`**（本 ckpt） | **0.0808** ✨ | **0.1515** ✨ | **10** | **16** |
+| `qwen3-30b-a3b-OpenClaw-RL-iter351` (later) | 0.0556 ⚠ | 0.0758 ⚠ | 5 | 11 |
+| `qwen3-30b-a3b-OpenClaw-RL-iter367` (final) | 0.0657 | 0.1212 | 8 | 13 |
+
+**iter_335 是这次 fresh run 中唯一 pass@1 / pass@3 / unique_solved 三项全部超过 base 的 ckpt**。iter_351 在中间出现明显回落（甚至跌穿 base），iter_367 部分恢复但仍未回到 iter_335 峰。详细分析见 [issue #14 eval comment](https://github.com/HansBug/OpenClaw-RL/issues/14#issuecomment-4537976268)。
+
+### vs 8B run-3（[issue #8](https://github.com/HansBug/OpenClaw-RL/issues/8) 同 66-task 标尺）
+
+![leaderboard](leaderboard.png)
+
+| Model | pass@1 | pass@3 | 体量 | 备注 |
+|---|---:|---:|---|---|
+| **Qwen3-30B-A3B-OpenClaw-RL-iter335**（本 ckpt） | **0.081** | **0.152** | 30B / 3B active MoE | **本 ckpt** |
+| Qwen3-30B-A3B (vanilla base) | 0.061 | 0.121 | 30B / 3B active | 训练起点 |
+| Qwen3-8B-OpenClaw-RL-iter215 ([issue #8](https://github.com/HansBug/OpenClaw-RL/issues/8)) | 0.056 | 0.091 | 8B dense | 8B run-3 OOD 峰 |
+| Qwen3-8B (base) | 0.020 | 0.030 | 8B dense | 8B 起点 |
+| Qwen3-235B-A22B + Terminus 1 | 0.066 | n/a | 235B / 22B active | TB 官方 80-task |
+| DeepSeek-R1 + Terminus 1 | 0.057 | n/a | 671B / 37B active | TB 官方 80-task |
+| Qwen3-32B + TerminalAgent | 0.155 | n/a | 32B dense | TB 官方 80-task |
+| Claude 4.5 Sonnet | 0.645 | n/a | 闭源 | frontier |
+| GPT-5 | 0.525 | n/a | 闭源 | frontier |
+
+> 注：本 ckpt 跑的是 66-task 子集（[harbor migration 失败 20 task](https://github.com/HansBug/OpenClaw-RL/issues/8) 后剩下的可跑部分），与 leaderboard 80-task 全集**不可直接横比**；但唯一靠谱的同标尺对比是与 issue #8 同样 66 子集的 Qwen3-8B 数字，本 ckpt 的 pass@1 = 0.081 比 Qwen3-8B-iter215 (0.056) 高 **44%**。
+
+### 任务级别的 OOD 解题分布
+
+![per-task heatmap](per_task_heatmap.png)
+
+14 个 task 在所有 4 个 ckpt × 3 attempt 中至少被解出 1 次，其余 52 task 全部 0 solved（与 [issue #8](https://github.com/HansBug/OpenClaw-RL/issues/8) 60/66 全 0 同 pattern）。iter_335 独占解出的 task：`csv-to-parquet`、`path-tracing-reverse`、`processing-pipeline 2x`。
+
+---
+
+## 为什么是 iter_335 而不是最新 ckpt
+
+完整 30B fresh run 跑到 iter_367 停（rollout 367 ≈ 训练 4.5 天），按 retention daemon 配置仅保留 latest 3 ckpt：iter_335 / iter_351 / iter_367。所有 4 个候选（含 base）都在 TB v0.1.x 跑过 OOD eval，得到的训练→OOD 轨迹是：
+
+| ckpt | rollout | in-domain raw_reward (训练侧) | OOD pass@1 | 训练→OOD 是否对齐？ |
+|---|---:|---:|---:|---|
+| base | — | — | 0.0606 | — |
+| **iter_335** | **335** | **+0.21** | **0.0808** ✨ | ✅ 训练 ↑ + OOD ↑ |
+| iter_351 | 351 | +0.18（含全局峰 r346 +0.5651） | 0.0556 | ❌ 训练保持高位但 OOD 跌穿 base |
+| iter_367 | 367 | +0.232（最高窗口均值） | 0.0657 | ❌ 训练 ↑ 但 OOD 没跟上 |
+
+iter_335 是这次训练**唯一明确"训练侧 ↑ + OOD ↑"的 ckpt**。这是 [issue #8 iter215](https://github.com/HansBug/OpenClaw-RL/issues/8)（8B run-3 也在中段 ckpt 拿 OOD 峰，最后回落）同款现象。如果你要拿一个最有可能在新 OOD task 上 work 的 30B fresh-run ckpt，**应该用这个 iter_335**，不要用最新的 iter_367。
+
+---
+
+## 推理时的 on-the-wire 协议
+
+**训练时这个 ckpt 看到的协议是 [camel-ai `TerminalToolkit`](https://docs.camel-ai.org/key_modules/terminaltoolkit) 的 4 工具 OpenAI function-calling**（与 [`Qwen3-8B-OpenClaw-RL-iter215`](https://huggingface.co/HansBug/Qwen3-8B-OpenClaw-RL-iter215) 完全相同），**不是** terminal-bench 的 `terminus-2` JSON 协议。
+
+| 文件 | 说明 |
+|---|---|
+| [`terminal-rl/agent/camel_agent.py`](https://github.com/HansBug/OpenClaw-RL/blob/main/terminal-rl/agent/camel_agent.py) | rollout agent = `camel.agents.ChatAgent` 子类 |
+| [`terminal-rl/remote/terminal_env.py`](https://github.com/HansBug/OpenClaw-RL/blob/main/terminal-rl/remote/terminal_env.py) | env 在 reset 时把 `camel.toolkits.TerminalToolkit` 的 4 个工具（`shell_exec` / `shell_view` / `shell_write_to_process` / `shell_write_content_to_file`）封装成 OpenAI tool schema 喂给 rollout |
+| [`terminal-rl/configs/rollout_qwen3.yaml`](https://github.com/HansBug/OpenClaw-RL/blob/main/terminal-rl/configs/rollout_qwen3.yaml) | `tool_call_parser: qwen25` — sglang 把模型 `<tool_call>{...}</tool_call>` markup 转回 OpenAI tool_calls |
+
+System prompt 是 [`terminal-rl/agent/camel_agent.py::get_developer_agent_prompt()`](https://github.com/HansBug/OpenClaw-RL/blob/main/terminal-rl/agent/camel_agent.py) 在 `system='Linux (in Docker)', machine='x86_64', non_think_mode=True` 配置下产出，结尾加 `/no_think` 关闭 qwen3 的 think trace。
+
+如果你想 100% 复刻训练分布在 inference 用这个 ckpt，用 [`HansBug/oc-repl`](https://github.com/HansBug/oc-repl) 的默认 `--protocol camel-terminal-toolkit`。
+
+---
+
+## 已知限制
+
+1. **30B 在 OOD 上的 RL 边际收益比 8B 小**。本 ckpt vs 30B base：pass@1 +33%；[Qwen3-8B-iter215](https://huggingface.co/HansBug/Qwen3-8B-OpenClaw-RL-iter215) vs Qwen3-8B base 是 +180%。如果你的目标是 frontier OOD 性能，纯 RL 已经 saturating，需要换 cold-start SFT。这与 [issue #11 容量假说](https://github.com/HansBug/OpenClaw-RL/issues/11) 一致。
+
+2. **iter_351 / iter_367 的存在说明 RL 训练后期有 OOD 漂移风险**。iter_335 是这次训练全 4 ckpt 中唯一 net-positive 的；更晚的 iter_351 反而跌穿 base。如果你打算复现这个训练 pipeline，建议把 ckpt freeze 频率提高，每 50-100 rollout 都做 OOD eval 看 trajectory。
+
+3. **模型不会主动停止 tool_call**。训练时 agent loop 由 `max_iteration: 10` 硬截断 + outcome reward 兜底，inference 时要自己设上限（[oc-repl 默认 12 轮](https://github.com/HansBug/oc-repl/blob/main/src/oc_repl/engine.py)）。
+
+4. **复杂多步 task adherence 中等**。简单任务（chmod、cat、ls）adherence 满分；多文件 Python 服务器、复杂 awk 这种偶尔会 thinking 太久没产出 tool_call。
+
+5. **terminus-2 / terminus-XML 协议是 OOD**。模型没在这些协议上 RL 训练过，能不能跑通靠 qwen3 底座的通用指令跟随能力。要复刻训练分布请用 camel TerminalToolkit 协议。
+
+6. **不是 frontier 水平**。pass@1 0.081 跟 Qwen3-235B 同水位（甚至略高），但跟 Claude 4.5 / GPT-5 还有 6-8× 差距。这个 ckpt 适合做 RL 训练框架的 baseline / case study / MoE base 上 RL 的对比基准，不适合直接当 production agent 用。
+
+---
+
+## 复现 / 推理工具
+
+| 仓库 / 工具 | 用途 |
+|---|---|
+| [`HansBug/OpenClaw-RL`](https://github.com/HansBug/OpenClaw-RL) | 完整的训练框架（slime + Megatron + sglang），含 launch script、配置、agent code |
+| [`HansBug/oc-repl`](https://github.com/HansBug/oc-repl) | 针对这个 family ckpt 的 Codex 风格 REPL，支持 4 种推理协议 (`camel-terminal-toolkit` 默认 = 训练分布字节级复刻) |
+| [`HansBug/Qwen3-8B-OpenClaw-RL-iter215`](https://huggingface.co/HansBug/Qwen3-8B-OpenClaw-RL-iter215) | **同 family** 的 8B ckpt（base = Qwen3-8B），用于 same-protocol 同标尺横比 |
+| [`HansBug/Qwen3-8B-OpenClaw-RL-tboverfit-iter311`](https://huggingface.co/HansBug/Qwen3-8B-OpenClaw-RL-tboverfit-iter311) | 同 family 的 8B eval-as-train 上界 probe ckpt |
+| [`HansBug/OpenClaw-RL` issue #14](https://github.com/HansBug/OpenClaw-RL/issues/14) | 本 ckpt 的完整训练 + eval 报告（118h 训练曲线 + 4 ckpt OOD eval + vs 8B 对比） |
+| [`HansBug/OpenClaw-RL` issue #4](https://github.com/HansBug/OpenClaw-RL/issues/4) | 8B run-3 训练报告（同算法 / 同 dataset / 同 lr，仅 base 模型不同） |
+| [`HansBug/OpenClaw-RL` issue #8](https://github.com/HansBug/OpenClaw-RL/issues/8) | 8B run-3 在 TB v0.1.x 66-task 的同标尺 OOD eval（本 ckpt eval 完全复用了该协议） |
+
+---
+
+## 引用 / 致谢
+
+- 基座：[`Qwen/Qwen3-30B-A3B`](https://huggingface.co/Qwen/Qwen3-30B-A3B)（Apache-2.0，vanilla 版，2025-04 release）
+- 训练框架：[`HansBug/OpenClaw-RL`](https://github.com/HansBug/OpenClaw-RL)
+- Agent harness：[`camel-ai/camel`](https://github.com/camel-ai/camel) [`TerminalToolkit`](https://docs.camel-ai.org/key_modules/terminaltoolkit)
+- Eval benchmark：[`laude-institute/terminal-bench`](https://github.com/laude-institute/terminal-bench)
+- Rollout 训练数据 pool：[`camel-ai/seta`](https://github.com/camel-ai/seta) 1376 task
+
+License: Apache-2.0（继承自 Qwen3-30B-A3B）。
+
+如果你用这个 ckpt 做 paper / blog post，欢迎引用 [`HansBug/OpenClaw-RL`](https://github.com/HansBug/OpenClaw-RL) 仓库 + 这个 model card。
--- a/config.json
+++ b/config.json
@@ -0,0 +1,38 @@
+{
+  "architectures": [
+    "Qwen3MoeForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "decoder_sparse_step": 1,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 48,
+  "mlp_only_layers": [],
+  "model_type": "qwen3_moe",
+  "moe_intermediate_size": 768,
+  "norm_topk_prob": true,
+  "num_attention_heads": 32,
+  "num_experts": 128,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 48,
+  "num_key_value_heads": 4,
+  "output_router_logits": false,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "router_aux_loss_coef": 0.001,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}
--- a/eval_pass_rate.png
+++ b/eval_pass_rate.png
--- a/generation_config.json
+++ b/generation_config.json
@@ -0,0 +1,13 @@
+{
+    "bos_token_id": 151643,
+    "do_sample": true,
+    "eos_token_id": [
+        151645,
+        151643
+    ],
+    "pad_token_id": 151643,
+    "temperature": 0.6,
+    "top_k": 20,
+    "top_p": 0.95,
+    "transformers_version": "4.51.0"
+}
--- a/leaderboard.png
+++ b/leaderboard.png
--- a/merges.txt
+++ b/merges.txt
--- a/model-00000-of-00012.safetensors
+++ b/model-00000-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d87654f409ae0c2daf1bbf8d52aad7796f973de28e39385dab5bece8f80df646
+size 5368403824
--- a/model-00001-of-00012.safetensors
+++ b/model-00001-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ace3b84fae9d386cde1bfeafbff6643ab4821db2697f6c04ac2126e8b1988a6
+size 5366338784
--- a/model-00002-of-00012.safetensors
+++ b/model-00002-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcad34d518774682dcd0cc69ef11961547872e2abe6457d9fcae85a25f288b8e
+size 5365807096
--- a/model-00003-of-00012.safetensors
+++ b/model-00003-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e1e77f2cb4f880cb55e383d24bf8946e8704c25be3ccbccb9f89007c083cfca
+size 5365807912
--- a/model-00004-of-00012.safetensors
+++ b/model-00004-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:773de90228818b49c2e90107c7f571a8a6bb4352640cfcec34de7f2340a9ec97
+size 5366340528
--- a/model-00005-of-00012.safetensors
+++ b/model-00005-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3dd1923ee40bceda9c58cd1f4c29af7fe9bd5c64e059915e2189dbd8e532ce0
+size 5365807832
--- a/model-00006-of-00012.safetensors
+++ b/model-00006-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e44b393301da6a6e77ea28bd5edacbb8a5255999f1fabb828b5b71c823274f0
+size 5365807888
--- a/model-00007-of-00012.safetensors
+++ b/model-00007-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:034830e720f81bd8756538629da33b20b377f64a4463205c06028d040819b771
+size 5365807968
--- a/model-00008-of-00012.safetensors
+++ b/model-00008-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddf4a72a4efd9bea144694b509598a312280e5ada55dedd5b2d54611bc38c50d
+size 5366340408
--- a/model-00009-of-00012.safetensors
+++ b/model-00009-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26e9bb6013595576fe90d0316b3a33d4ccfb8d02acb4c39d3c3d9b0cb65aba89
+size 5365807856
--- a/model-00010-of-00012.safetensors
+++ b/model-00010-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad77a88f4bc90b14112bc51315360faa6c1dd75ffa9fe75339abeefaadfa95b1
+size 5365807960
--- a/model-00011-of-00012.safetensors
+++ b/model-00011-of-00012.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0727dee24c5c05100124a647c1fb746a041c1a95db56c916e5b32394a2c921c
+size 2038500104
--- a/model.safetensors.index.json
+++ b/model.safetensors.index.json
--- a/per_task_heatmap.png
+++ b/per_task_heatmap.png
--- a/tokenizer.json
+++ b/tokenizer.json
--- a/tokenizer_config.json
+++ b/tokenizer_config.json
@@ -0,0 +1,239 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if message.content is string %}\n        {%- set content = message.content %}\n    {%- else %}\n        {%- set content = '' %}\n    {%- endif %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- endif %}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}
--- a/training_curves.png
+++ b/training_curves.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae8651c73c689c25cbf5ea9014c0e9673ecaac6d5463ad32ec173574d0485398
+size 391498
--- a/vocab.json
+++ b/vocab.json