Upload Valen-4B merged model and unmerged LoRA bundle

Final shared-state Qwen3.5-4B + two-layer Mixer after 1225k two-stage SFT (step 325). Includes standalone merged weights, original LoRA/Mixer/visual-merger weights, processor, tokenizer, custom loading code and export validation reports.

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* 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
+unmerged/tokenizer.json filter=lfs diff=lfs merge=lfs -text

COMPLETE.json

@@ -0,0 +1,6 @@
+{
+  "model": "Valen-4B",
+  "validated": true,
+  "files": 21,
+  "source_checkpoint_sha256": "b4193b4e907f44b913fd54a8d6b998e7050a1528306269a000f1e5e8bd8dffc7"
+}

LICENSE

@@ -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 2026 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.

MODEL_USAGE.md

@@ -0,0 +1,49 @@
+# Valen-4B
+
+完整合并的 Qwen3.5 + 两层 Mixer，提供 Choice、Noul、Score 决策和 share_state 多题推理。
+Complete Qwen3.5 backbone with merged LoRA, trained visual merger, and the FP32 Mixer head.
+
+训练版本：1195k + JevBench 30k，两阶段 SFT，最终 step 325。
+全体基座参数、训练参数、分词器及输入编译代码均在本目录。
+
+## 加载 / Load
+
+验证环境：Python 3.10、PyTorch 2.6.0、Transformers 5.4.0、torchvision 0.21.0、Pillow、av 16.1.0。
+使用 Flash Attention 2 时，另需与 PyTorch/CUDA 匹配的 flash-attn。
+
+```python
+import torch
+from transformers import AutoModel
+
+model = AutoModel.from_pretrained(
+    "./Valen-4B",
+    trust_remote_code=True,
+    dtype=torch.bfloat16,
+    attn_implementation="sdpa",  # 或 flash_attention_2 / or flash_attention_2
+).to("cuda").eval()
+# 和导出一致性验证保持相同的 FP32 精度。 / Match export verification precision.
+torch.set_float32_matmul_precision("highest")
+torch.backends.cudnn.allow_tf32 = False
+
+response = model.predict({
+    "state": "There is one cat in the room.",
+    "questions": {
+        "animal": {"type": "choice", "instructions": "Which animal is present?",
+                   "criteria": {"cat": "A cat", "dog": "A dog"}},
+        "is_cat": {"type": "noul", "instructions": "There is a cat in the room."},
+        "count": {"type": "score", "instructions": "How many cats are present?",
+                  "criteria": ["No cats", "Exactly one cat", "Two or more cats"]},
+    },
+})
+print(response["answers"])
+```
+
+图像/视频使用 Valen 的 messages、image_url、video_url 输入格式，媒体路径由 `media_root` 解析。
+默认 `execution="shared_state"`；`model.predict(..., execution="question")` 可使用独立问题路径。
+视频沿用训练配置，采样 16 帧。所有 Mixer 参数保持 FP32，基座权重为 BF16。
+
+`validation.json` 记录合并前后概率差异、重新加载一致性、文本/图像/视频及三类题型检查。
+`export_manifest.json` 记录来源 checkpoint、基座版本及文件 SHA-256。
+LoRA 以 FP32 累加后保存为 BF16；融合矩阵的舍入顺序与独立 LoRA 分支不同。
+验证样本中最大概率绝对差：0.01969546；预测标签变化：0 / 20。
+该检查只验证导出样本，不能代替完整 benchmark；之前 checkpoint 的成绩属于未融合版本。

README.md

@@ -1,3 +1,106 @@
 ---
 license: apache-2.0
+base_model: Qwen/Qwen3.5-4B
+library_name: transformers
+tags:
+- valen
+- decision-making
+- shared-state
+- qwen3.5
+- custom-code
 ---
+
+# Valen-4B
+
+Valen-4B combines Qwen3.5-4B with a two-layer MLP-Mixer decision head. It predicts probability distributions for Choice, Noul (yes/no), and Score questions. A shared state can contain text, images, or video and serve multiple questions.
+
+Valen-4B 基于 Qwen3.5-4B 和两层 MLP-Mixer 决策头，支持图文、文本、视频，以及共享 state 的多题推理。
+
+## Training / 训练
+
+- Two-stage SFT: decision-head warmup followed by joint LoRA, Mixer, and visual-merger fine-tuning.
+- Training data: the previous 1,195,000 records plus JevBench train_v2 30,000 records, totaling 1,225,000 records. This count refers to records; shared-state records may contain multiple questions.
+- Warmup: 122,500 records, 67 optimizer steps. Joint SFT: one epoch over all 1,225,000 records, 325 optimizer steps, on 32 H200 GPUs.
+- Final checkpoint: joint SFT step 325. This release uses SFT training.
+- Backbone: BF16. Mixer: FP32. Videos use 16 sampled frames.
+
+## Files / 权重格式
+
+- Root directory: complete backbone with merged LoRA, trained visual merger, FP32 Mixer, tokenizer, processor, and standalone loading code.
+- `unmerged/`: separate PEFT LoRA, Mixer, and trained visual-merger weights, with its own configuration and loading code. Requires the pinned original Qwen3.5-4B base.
+- `MODEL_USAGE.md`: inference examples and environment details.
+- `validation.json`, `fresh_process_validation.json`: export checks covering text, images, video, Choice/Noul/Score, question/shared_state, and isolated Python loading.
+- `export_manifest.json`: source revision and file checksums.
+
+## Load merged weights / 加载完整合并权重
+
+The verified environment uses Python 3.10, PyTorch 2.6.0, Transformers 5.4.0, torchvision 0.21.0, Pillow, and av 16.1.0. Flash Attention 2 requires a compatible flash-attn installation. Private repositories require an account with access.
+
+```python
+import torch
+from transformers import AutoModel
+
+model = AutoModel.from_pretrained(
+    "Valen-Team/Valen-4B",
+    trust_remote_code=True,
+    dtype=torch.bfloat16,
+    attn_implementation="sdpa",  # or "flash_attention_2"
+).to("cuda").eval()
+torch.set_float32_matmul_precision("highest")
+torch.backends.cudnn.allow_tf32 = False
+
+response = model.predict({
+    "state": "There is one cat in the room.",
+    "questions": {
+        "animal": {"type": "choice", "instructions": "Which animal is present?",
+                   "criteria": {"cat": "A cat", "dog": "A dog"}},
+        "is_cat": {"type": "noul", "instructions": "There is a cat in the room."},
+    },
+}, execution="shared_state")
+print(response["answers"])
+```
+
+For a local copy, replace the repository ID with `"./Valen-4B"`.
+`execution="question"` scores questions separately. Images and video use the Valen `messages`, `image_url`, and `video_url` schema; relative media paths are resolved against `media_root`.
+
+## Load unmerged LoRA / 加载未合并权重
+
+Download the repository first; the adapter loader expects a local directory. PEFT 0.18.1 and safetensors are also required.
+
+```python
+import torch
+from huggingface_hub import snapshot_download
+from transformers import AutoModel
+
+folder = snapshot_download("Valen-Team/Valen-4B")
+model = AutoModel.from_pretrained(
+    folder + "/unmerged",
+    trust_remote_code=True,
+    local_files_only=False,  # download the pinned Qwen base if needed
+    dtype=torch.bfloat16,
+    attn_implementation="sdpa",
+).to("cuda").eval()
+torch.set_float32_matmul_precision("highest")
+torch.backends.cudnn.allow_tf32 = False
+```
+
+For an existing local base, pass `base_model_path="./Qwen3.5-4B"` and `local_files_only=True`. The loader verifies its original weight hashes and restores LoRA, Mixer, and visual-merger weights together. `is_trainable=True` enables training those parameters.
+
+## Reference benchmark results / 参考评测
+
+Measured on the original final **unmerged** SFT checkpoint, using shared-state inference for multimodal sets and single-question inference for JevBench:
+
+| Benchmark | QA | Accuracy | Choice | Noul | Score |
+| --- | --- | --- | --- | --- | --- |
+| eval_v1 | 2,000 | 83.06% | 85.42% | 83.25% | 54.00% |
+| General | 10,000 | 88.68% | 90.68% | 85.76% | 75.60% |
+| Video | 7,143 | 87.36% | 90.80% | 93.14% | 72.43% |
+| JevBench | 231 | 83.12% | 84.89% | 81.08% | 77.78% |
+
+JevBench difficulty scores: Easy 100.00% (48/48), Standard 93.06% (67/72), Hard 69.37% (77/111). The overall score is weighted by task count: 192/231.
+
+The export smoke cases showed a maximum absolute probability difference of 0.01969546 between merged and unmerged inference, with no label changes in those cases. This does not establish full-benchmark equivalence.
+
+For eval_v1, accuracy uses 1,700 hard-label questions. Its remaining 300 soft-label Score questions contribute probability metrics and are excluded from hard-label accuracy.
+
+Merged BF16 matrices and separate LoRA branches can produce numerical differences. Export validation records those differences; the merged release has not been independently rerun on the complete benchmarks. The unmerged bundle preserves the original trained tensors and passed reloading checks.

batching.py

@@ -0,0 +1,62 @@
+"""Pad language rows and concatenate visual patches. / 文本补齐，视觉 patch 拼接。"""
+import torch
+from torch.nn.utils.rnn import pad_sequence
+
+
+SEQUENCE_KEYS = {"input_ids", "attention_mask", "mm_token_type_ids"}
+VISUAL_KEYS = {"pixel_values", "pixel_values_videos", "image_grid_thw", "video_grid_thw"}
+
+
+def collate_branches(branches, pad_token_id=0):
+    if not branches:
+        raise ValueError("Cannot collate an empty branch batch")
+    inputs = [branch.inputs for branch in branches]
+    unknown = set().union(*(row.keys() for row in inputs)) - SEQUENCE_KEYS - VISUAL_KEYS
+    if unknown:
+        raise ValueError(f"Unsupported batched Qwen inputs: {sorted(unknown)}")
+    if any(row["input_ids"].ndim != 2 or row["input_ids"].shape[0] != 1 for row in inputs):
+        raise ValueError("Each compiled branch must contain one language row")
+    result = {}
+    for key in SEQUENCE_KEYS:
+        if key != "input_ids" and not any(key in row for row in inputs):
+            continue
+        rows = []
+        for row in inputs:
+            value = row.get(key)
+            if value is None:
+                value = (torch.ones_like(row["input_ids"]) if key == "attention_mask"
+                         else torch.zeros_like(row["input_ids"]))
+            if value.shape != row["input_ids"].shape:
+                raise ValueError(f"{key} must match the branch input_ids shape")
+            rows.append(value[0])
+        # Right padding keeps all compiler readout positions and spans unchanged.
+        # 右侧补齐，候选位置和角色区间无需偏移。
+        result[key] = pad_sequence(rows, batch_first=True,
+                                   padding_value=pad_token_id if key == "input_ids" else 0)
+    if "attention_mask" not in result:
+        result["attention_mask"] = pad_sequence([torch.ones_like(r["input_ids"][0]) for r in inputs],
+                                                batch_first=True, padding_value=0)
+    for key in VISUAL_KEYS:
+        values = [row[key] for row in inputs if key in row]
+        if values:
+            result[key] = torch.cat(values, dim=0)
+    return result
+
+
+def branch_batches(questions, max_tokens):
+    """Limit padded tokens per backbone forward, keeping Score branch order."""
+    if type(max_tokens) is not int or max_tokens <= 0:
+        raise ValueError("microbatch_max_tokens must be a positive integer")
+    batch, longest = [], 0
+    for question_index, question in enumerate(questions):
+        if not question.branches:
+            raise ValueError("A Qwen question requires at least one branch")
+        for branch in question.branches:
+            length = branch.inputs["input_ids"].shape[1]
+            if batch and max(longest, length) * (len(batch) + 1) > max_tokens:
+                yield batch
+                batch, longest = [], 0
+            batch.append((question_index, branch))
+            longest = max(longest, length)
+    if batch:
+        yield batch

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

compiler.py

@@ -0,0 +1,239 @@
+"""Compile Qwen question branches or a shared sequence with per-question readouts."""
+from copy import deepcopy
+from dataclasses import dataclass
+from pathlib import Path
+import hashlib
+import torch
+from .schema import candidates, target_distribution
+from .data_types import QuestionSpec
+
+
+@dataclass
+class Branch:
+    inputs: dict
+    candidate_positions: list
+    decision_position: int
+    context_span: tuple | None = None
+    instruction_span: tuple | None = None
+    candidate_spans: list | None = None
+
+
+@dataclass(init=False)
+class Question(QuestionSpec):
+    branches: list
+
+    def __init__(self, qid, kind, keys, descriptions, branches, target, is_text):
+        # Keep the positional constructor used by existing compiler clients.
+        super().__init__(qid, kind, keys, descriptions, target, is_text)
+        self.branches = branches
+
+
+@dataclass
+class CompiledState:
+    questions: list
+    logical_tokens: int
+    compute_tokens: int
+    media: list
+    inputs: dict | None = None
+
+
+class Compiler:
+    def __init__(self, processor, media_root=".", max_length=8192, media_kwargs=None, verified_media_hashes=None,
+                 execution="question"):
+        if execution not in {"question", "shared_state"}:
+            raise ValueError(f"Unknown qwen_execution: {execution}")
+        self.execution = execution
+        self.processor = processor
+        self.tokenizer = processor.tokenizer
+        self.media_root = Path(media_root)
+        self.max_length = max_length
+        self.media_kwargs = media_kwargs or {}
+        # Optional, pre-audited immutable assets for latency measurements.
+        # Normal training/inference continues to hash each input at use time.
+        self.verified_media_hashes = verified_media_hashes or {}
+
+    def _messages(self, state):
+        messages = [{"role": "user", "content": state}] if isinstance(state, str) else deepcopy(state["messages"])
+        media = []
+        for message in messages:
+            if message["role"] not in {"user", "assistant", "system"}:
+                raise ValueError("Unsupported state role")
+            if isinstance(message["content"], str):
+                message["content"] = [{"type": "text", "text": message["content"]}]
+            for item in message["content"]:
+                kind = item["type"]
+                if kind == "text":
+                    # Prevent user text from injecting Qwen control tokens.
+                    if any(t in item["text"] for t in self.tokenizer.all_special_tokens):
+                        raise ValueError("Reserved tokenizer control token in state")
+                    continue
+                if kind not in {"image_url", "video_url"}:
+                    raise ValueError(f"Unsupported content type: {kind}")
+                url = item[kind]["url"]
+                if "://" in url:
+                    raise ValueError("Materialize remote media locally with a hash before training")
+                path = Path(url)
+                path = path if path.is_absolute() else self.media_root / path
+                path = path.resolve(strict=True)
+                digest = self.verified_media_hashes.get(str(path))
+                if digest is None:
+                    digest = hashlib.sha256(path.read_bytes()).hexdigest()
+                media.append({"path": str(path), "sha256": digest, "type": kind})
+                item.clear()
+                item.update(type=kind.removesuffix("_url"), path=str(path))
+        return messages, media
+
+    def compile(self, record, rng=None, labeled_only=False):
+        selected = []
+        for qid, question in record["request"]["questions"].items():
+            pairs = candidates(question)
+            target = record.get("targets", {}).get(qid)
+            if labeled_only and self.execution == "question" and target_distribution(target, [k for k, _ in pairs]) is None:
+                continue
+            if rng is not None and question["type"] == "choice":
+                rng.shuffle(pairs)
+            selected.append((qid, question, pairs, target))
+        # 标签只决定监督范围，共享序列仍保留所有题。 / Labels select losses, never prompt content.
+        if not selected or (labeled_only and not any(
+                target_distribution(target, [k for k, _ in pairs]) is not None
+                for _, _, pairs, target in selected)):
+            return CompiledState([], 0, 0, [])
+        messages, media = self._messages(record["request"]["state"])
+        expected_assets = {str((self.media_root / a["path"]).resolve()): a["sha256"] for a in record.get("assets", [])}
+        for entry in media:
+            if entry["path"] in expected_assets and expected_assets[entry["path"]] != entry["sha256"]:
+                raise ValueError(f"Media hash mismatch: {entry['path']}")
+        base = dict(self.processor.apply_chat_template(
+            messages, tokenize=True, add_generation_prompt=False, return_dict=True,
+            return_tensors="pt", processor_kwargs=dict(self.media_kwargs, return_metadata=True)))
+        # Metadata is audit information, never a backbone keyword argument.
+        metadata = base.pop("video_metadata", None)
+        if metadata is not None:
+            for entry, meta in zip([m for m in media if m["type"] == "video_url"], metadata):
+                entry["sampling"] = dict(meta) if isinstance(meta, dict) else vars(meta)
+                sampling = entry["sampling"]
+                indices = sampling.get("frames_indices")
+                if indices is not None and sampling.get("fps"):
+                    sampling["frames_indices"] = [int(i) for i in indices]
+                    sampling["timestamps_seconds"] = [int(i) / sampling["fps"] for i in indices]
+                    sampling["visible_range_seconds"] = [min(sampling["timestamps_seconds"]), max(sampling["timestamps_seconds"])]
+        for name in ("image_grid_thw", "video_grid_thw"):
+            if name in base:
+                grid = base[name]
+                media.append({name: grid.tolist(), "visual_tokens": int(grid.prod(-1).sum()) // 4})
+        base_length = base["input_ids"].shape[1]
+        if self.execution == "shared_state":
+            return self._compile_shared(selected, base, media, labeled_only)
+        logical_tokens, compute_tokens = base_length, 0
+        compiled = []
+        for qid, question, pairs, target in selected:
+            keys = [k for k, _ in pairs]
+            branches = []
+            groups = [[pair] for pair in pairs] if question["type"] == "score" else [pairs]
+            for group in groups:
+                suffix = []
+                def append(text, char_span=None):
+                    start = base_length + len(suffix)
+                    if char_span is None:
+                        suffix.extend(self.tokenizer.encode(text, add_special_tokens=False))
+                        return None
+                    # 保持原有分词，只额外记录区间。 / Preserve token IDs; record role spans.
+                    encoded = self.tokenizer(text, add_special_tokens=False, return_offsets_mapping=True)
+                    suffix.extend(encoded["input_ids"])
+                    indices = [i for i, (a, b) in enumerate(encoded["offset_mapping"])
+                               if a < char_span[1] and b > char_span[0]]
+                    if not indices:
+                        raise ValueError("Empty token span for decision role")
+                    return (start + indices[0], start + indices[-1] + 1)
+                # Canonical segment tokenization makes endpoints exact; no string searching.
+                suffix.extend(self.tokenizer.encode("<|im_start|>user\n", add_special_tokens=False))
+                for value in [question["instructions"]] + [v for pair in group for v in pair]:
+                    if any(t in value for t in self.tokenizer.all_special_tokens):
+                        raise ValueError("Reserved tokenizer control token in question/criteria")
+                prefix = "Task: " + question["type"] + "\nQuestion: "
+                instruction_span = append(prefix + question["instructions"] + "\nCandidates:\n",
+                                          (len(prefix), len(prefix) + len(question["instructions"])))
+                positions, candidate_spans = [], []
+                for key, description in group:
+                    # Score never sees its level ID, nor any other level description.
+                    prefix = key + ": " if question["type"] != "score" else ""
+                    candidate_spans.append(append(prefix + description, (len(prefix), len(prefix) + len(description))))
+                    positions.append(base_length + len(suffix) - 1)
+                    append("\n")
+                append("Decision:")
+                length = base_length + len(suffix)
+                if length > self.max_length:
+                    raise ValueError(f"{qid}: {length} tokens exceed max_length={self.max_length}; no truncation")
+                inputs = dict(base)
+                inputs["input_ids"] = torch.cat([base["input_ids"], torch.tensor([suffix], dtype=torch.long)], dim=1)
+                inputs["attention_mask"] = torch.ones_like(inputs["input_ids"])
+                if "mm_token_type_ids" in base:
+                    inputs["mm_token_type_ids"] = torch.cat([base["mm_token_type_ids"], torch.zeros((1, len(suffix)), dtype=torch.long)], dim=1)
+                branches.append(Branch(inputs, positions, length - 1, (0, base_length),
+                                       instruction_span, candidate_spans))
+                compute_tokens += length
+                logical_tokens += len(suffix)
+            compiled.append(Question(qid, question["type"], keys, [v for _, v in pairs], branches,
+                                     target_distribution(target, keys), not media))
+        return CompiledState(compiled, logical_tokens, compute_tokens, media)
+
+    def _compile_shared(self, selected, base, media, labeled_only=False):
+        """All schemas precede all readouts; labels and external IDs never enter inputs.
+
+        This is ordinary causal attention, not isolated question/Score branches.
+        No answer tokens or LM output layer are required by the decision heads.
+        """
+        base_length = base["input_ids"].shape[1]
+        suffix, compiled = [], []
+
+        def append(text, span=None):
+            start = base_length + len(suffix)
+            if span is None:
+                suffix.extend(self.tokenizer.encode(text, add_special_tokens=False))
+                return None
+            encoded = self.tokenizer(text, add_special_tokens=False, return_offsets_mapping=True)
+            suffix.extend(encoded["input_ids"])
+            indices = [i for i, (a, b) in enumerate(encoded["offset_mapping"]) if a < span[1] and b > span[0]]
+            if not indices:
+                raise ValueError("Empty token span for decision role")
+            return start + indices[0], start + indices[-1] + 1
+
+        append("<|im_start|>user\n")
+        for index, (qid, question, pairs, target) in enumerate(selected):
+            for value in [question["instructions"]] + [v for pair in pairs for v in pair]:
+                if any(t in value for t in self.tokenizer.all_special_tokens):
+                    raise ValueError("Reserved tokenizer control token in question/criteria")
+            prefix = f"Question {index + 1}\nTask: {question['type']}\nQuestion: "
+            instruction = append(prefix + question["instructions"] + "\nCandidates:\n",
+                                 (len(prefix), len(prefix) + len(question["instructions"])))
+            positions, spans = [], []
+            for key, description in pairs:
+                prefix = key + ": " if question["type"] != "score" else ""
+                spans.append(append(prefix + description, (len(prefix), len(prefix) + len(description))))
+                positions.append(base_length + len(suffix) - 1)
+                append("\n")
+            # Inputs and the decision position are filled after all schemas are appended.
+            branch = Branch({}, positions, -1, (0, base_length), instruction, spans)
+            keys = [key for key, _ in pairs]
+            compiled.append(Question(qid, question["type"], keys, [v for _, v in pairs], [branch],
+                                     target_distribution(target, keys), not media))
+        append("<|im_end|>\n<|im_start|>assistant\n")
+        for index, question in enumerate(compiled):
+            append(f"Question {index + 1} Decision:")
+            question.branches[0].decision_position = base_length + len(suffix) - 1
+            append("\n")
+        length = base_length + len(suffix)
+        if length > self.max_length:
+            raise ValueError(f"Shared state: {length} tokens exceed max_length={self.max_length}; no truncation")
+        inputs = dict(base)
+        inputs["input_ids"] = torch.cat([base["input_ids"], base["input_ids"].new_tensor([suffix])], dim=1)
+        inputs["attention_mask"] = torch.cat([base["attention_mask"],
+                                             base["attention_mask"].new_ones((1, len(suffix)))], dim=1)
+        if "mm_token_type_ids" in base:
+            inputs["mm_token_type_ids"] = torch.cat([base["mm_token_type_ids"],
+                                                    base["mm_token_type_ids"].new_zeros((1, len(suffix)))], dim=1)
+        for question in compiled:
+            question.branches[0].inputs = inputs
+        if labeled_only:
+            compiled = [question for question in compiled if question.target is not None]
+        return CompiledState(compiled, length, length, media, inputs)

config.json

@@ -0,0 +1,138 @@
+{
+  "architectures": [
+    "ValenQwenForDecisionMaking"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_valen.ValenQwenConfig",
+    "AutoModel": "modeling_valen.ValenQwenForDecisionMaking"
+  },
+  "dtype": "bfloat16",
+  "image_token_id": 248056,
+  "model_type": "valen_qwen",
+  "qwen_execution": "shared_state",
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attn_output_gate": true,
+    "bos_token_id": null,
+    "dtype": "bfloat16",
+    "eos_token_id": 248044,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 2560,
+    "initializer_range": 0.02,
+    "intermediate_size": 9216,
+    "layer_types": [
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention"
+    ],
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 32,
+    "linear_value_head_dim": 128,
+    "mamba_ssm_dtype": "float32",
+    "max_position_embeddings": 262144,
+    "mlp_only_layers": [],
+    "model_type": "qwen3_5_text",
+    "mtp_num_hidden_layers": 1,
+    "mtp_use_dedicated_embeddings": false,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 4,
+    "pad_token_id": null,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        11,
+        11,
+        10
+      ],
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 10000000,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 248320
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.4.0",
+  "valen_head": {
+    "head_channel_hidden_dim": 1024,
+    "head_layers": 2,
+    "head_token_hidden_dim": 16,
+    "head_type": "mixer",
+    "head_width": 256
+  },
+  "valen_max_length": 32768,
+  "valen_media_kwargs": {
+    "images_kwargs": {
+      "size": {
+        "longest_edge": 262144,
+        "shortest_edge": 65536
+      }
+    },
+    "videos_kwargs": {
+      "fps": null,
+      "num_frames": 16,
+      "size": {
+        "longest_edge": 2097152,
+        "shortest_edge": 131072
+      }
+    }
+  },
+  "video_token_id": 248057,
+  "vision_config": {
+    "deepstack_visual_indexes": [],
+    "depth": 24,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1024,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4096,
+    "model_type": "qwen3_5",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 2560,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 248054,
+  "vision_start_token_id": 248053
+}

configuration_valen.py

@@ -0,0 +1,16 @@
+"""Standalone Qwen decision configuration. / 独立的 Qwen 决策模型配置。"""
+from transformers import Qwen3_5Config
+
+
+class ValenQwenConfig(Qwen3_5Config):
+    model_type = "valen_qwen"
+
+    def __init__(self, valen_head=None, qwen_execution="shared_state", valen_max_length=32768,
+                 valen_media_kwargs=None, **kwargs):
+        if qwen_execution not in {"question", "shared_state"}:
+            raise ValueError("qwen_execution must be question or shared_state")
+        self.valen_head = valen_head or {"head_type": "mixer"}
+        self.qwen_execution = qwen_execution
+        self.valen_max_length = valen_max_length
+        self.valen_media_kwargs = valen_media_kwargs or {}
+        super().__init__(**kwargs)

data_types.py

@@ -0,0 +1,12 @@
+"""Model-independent decision metadata; encoded tensors stay with each compiler."""
+from dataclasses import dataclass
+
+
+@dataclass
+class QuestionSpec:
+    qid: str
+    kind: str
+    keys: list
+    descriptions: list
+    target: list | None
+    is_text: bool

export_manifest.json

@@ -0,0 +1,82 @@
+{
+  "format": "valen_qwen_merged_v1",
+  "created_utc": "2026-10-06T06:52:07.182367+00:00",
+  "model": "Valen-4B",
+  "architecture": "Qwen3.5 + two-layer Mixer",
+  "execution": "shared_state",
+  "source_checkpoint_sha256": "b4193b4e907f44b913fd54a8d6b998e7050a1528306269a000f1e5e8bd8dffc7",
+  "training_step": 325,
+  "training_records": 1225000,
+  "training_data_sha256": "648cb127fe4830f9faaa6d4595445cc2573ef0c152f8f74c9c776dc8a9223d7f",
+  "base": {
+    "repo": "Qwen/Qwen3.5-4B",
+    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+    "weight_sha256": {
+      "model.safetensors-00001-of-00002.safetensors": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
+      "model.safetensors-00002-of-00002.safetensors": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188"
+    },
+    "files_sha256": {
+      ".gitattributes": "34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930",
+      "LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
+      "README.md": "1406be1b6b8fd8a6545870da516912804756593628a1d0fb0a7965211e82a7bb",
+      "chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+      "config.json": "ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670",
+      "merges.txt": "a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d",
+      "model.safetensors-00001-of-00002.safetensors": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
+      "model.safetensors-00002-of-00002.safetensors": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188",
+      "model.safetensors.index.json": "cf3f798ee02ba45f9622aa8892a47369ab667d0afbf154ee7c2212de42e6302d",
+      "preprocessor_config.json": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
+      "tokenizer.json": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42",
+      "tokenizer_config.json": "316230d6a809701f4db5ea8f8fc862bc3a6f3229c937c174e674ff3ca0a64ac8",
+      "video_preprocessor_config.json": "7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13",
+      "vocab.json": "ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003"
+    },
+    "verified_from": "download_manifest.json; every LFS SHA256 matched"
+  },
+  "merge": "FP32 base+LoRA accumulation, stored BF16; trained visual merger and FP32 Mixer retained",
+  "merged_linear_layers": 248,
+  "runtime_source_sha256": {
+    "valen/modeling/qwen/heads.py": "63b5970a8bfa121eff94d50ffd35e1fa8387f3af5abccc5f5b75381fb084fc34",
+    "valen/modeling/qwen/batching.py": "ea45d4724bcafca559fc5a2a18092918b633746771c19da6554c29bc5f2cef64",
+    "valen/modeling/qwen/model.py": "37b97e52bd16ada6a841fd553454ec0b0e458f3966ca0bef8a1bfe15293c2271",
+    "valen/data/compilers/qwen.py": "cae0389c407a9e67e1b14776b56d87767046b4ad119e93d138e1ae2f20fcadb2",
+    "valen/data/schema.py": "253825862b60016e37fd1040492008ad7f303679c092313d8a6527cf509f3b6e",
+    "valen/data/types.py": "b06ce08081d8f7b2bf36abf302410a96086395e75a04f9e9b1c591db519b1641",
+    "valen/evaluation/inference.py": "ccb47efd9f029f19a70cc462e262143cbab021deb4b8f4d11d2ecec054cbe553"
+  },
+  "versions": {
+    "torch": "2.6.0",
+    "transformers": "5.4.0",
+    "peft": "0.18.1",
+    "safetensors": "0.6.2",
+    "flash-attn": "2.7.4.post1",
+    "torchvision": "0.21.0",
+    "Pillow": "11.3.0",
+    "av": "16.1.0"
+  },
+  "files_sha256": {
+    "LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
+    "MODEL_USAGE.md": "cc6beb959e63059e68d8e5badb4df45090a22f1b049b1762262deb1988a61590",
+    "batching.py": "ea45d4724bcafca559fc5a2a18092918b633746771c19da6554c29bc5f2cef64",
+    "chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+    "compiler.py": "f35bab2e71b6beb1228daaa5b7baecdacf04eaf713aee14e8082b97090f910b4",
+    "config.json": "15dd1b0e94aa6ad330d3d8c50e4f9ea579edbb21fb2ba13c02b32c3e0ede6527",
+    "configuration_valen.py": "735b02402fb5b6c7a1652f6691ba66d066ef99e84500042faec396ff7a668fe5",
+    "data_types.py": "b06ce08081d8f7b2bf36abf302410a96086395e75a04f9e9b1c591db519b1641",
+    "heads.py": "63b5970a8bfa121eff94d50ffd35e1fa8387f3af5abccc5f5b75381fb084fc34",
+    "model-00001-of-00003.safetensors": "21fef5c5b93c2d9f1e54100f31709aaeb1651058784b9d38c47d9345fdba50d0",
+    "model-00002-of-00003.safetensors": "9ec6792a776db86b348b97ae3e9ff5feaa055dba2ea28035aac9f1d543623e9b",
+    "model-00003-of-00003.safetensors": "c7b96a0d85eb01475c0f775a51f88e4455aeae31143cc86c8457723d9019c13e",
+    "model.safetensors.index.json": "69f50ed2c7dcbbda30cc1b9ec3413ef84f2ac9aa569f357eec9f8ab63d1fdb95",
+    "modeling_valen.py": "07f87481334cb1e28e9daee036c27bab3654645abf3df7983346b316e7788848",
+    "processor_config.json": "d89ef49ce9cd37fbf510158e13c1ef063d9286411c1ec9049932dbe0487143b1",
+    "responses.py": "298016b0070f1000a38d033e1048e038ff5df6fc1dcbe830601105b181f1d228",
+    "runtime_model.py": "98bf0443ca196900b9f2b056129c84932041b8b53395d111f2fe6d3f2bc1e0a6",
+    "schema.py": "253825862b60016e37fd1040492008ad7f303679c092313d8a6527cf509f3b6e",
+    "tokenizer.json": "87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4",
+    "tokenizer_config.json": "e98f1901ac6f0adff67b1d540bfa0c36ac1a0cf59eb72ed78146ef89aafa1182",
+    "validation.json": "0ed2de1bc25b85c42d2329e42eb53bb6bad85e1ee3e2ec598a1dff643a3aa509",
+    "README.md": "fc0d6e3a980891edd2264d5a705da824cb1c0a31ebf52f44da54903c13d2b4c3",
+    "fresh_process_validation.json": "5945d5fc669cb573bb47db5fd41c3bb3ceb51b81b6951eea30681c7451505d43"
+  }
+}

fresh_process_validation.json

@@ -0,0 +1,112 @@
+{
+  "passed": true,
+  "isolated_python": true,
+  "valen_imports_blocked": true,
+  "all_real_modalities_and_question_types_passed": true,
+  "sdpa_example_passed": true,
+  "cases": 20,
+  "max_probability_abs_error": 0.0,
+  "max_logit_abs_error": 0.0,
+  "details": [
+    {
+      "case": "shared_state/text_multi/choice",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/text_multi/noul",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/text_multi/score",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/text_multi/score_many",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/image_score/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/image_noul/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/image_choice/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/video_multi/q000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/video_multi/score_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/video_multi/noul_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/choice",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/noul",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/score",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/score_many",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/image_score/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/image_noul/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/image_choice/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/video_multi/q000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/video_multi/score_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/video_multi/noul_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    }
+  ]
+}

heads.py

@@ -0,0 +1,143 @@
+"""可替换的候选评分头。 / Interchangeable, candidate-wise decision heads."""
+import math
+
+import torch
+from torch import nn
+
+
+def head_signature(config):
+    """用于构建与恢复校验。 / Canonical architecture settings for loading."""
+    kind = config.get("head_type", "bilinear")
+    defaults = {
+        "bilinear": {"projection_dim": 256},
+        "mlp": {"head_hidden_dim": 512, "head_bottleneck_dim": 128},
+        "role_mlp": {"head_width": 256, "head_hidden_dim": 1024},
+        "mixer": {"head_width": 256, "head_layers": 2,
+                  "head_token_hidden_dim": 16, "head_channel_hidden_dim": 1024},
+    }
+    if kind not in defaults:
+        raise ValueError(f"Unknown Qwen head_type: {kind!r}")
+    values = {key: config.get(key, value) for key, value in defaults[kind].items()}
+    if any(type(value) is not int or value <= 0 for value in values.values()):
+        raise ValueError("Decision head dimensions must be positive integers")
+    return {"head_type": kind, **values}
+
+
+class DecisionHead(nn.Module):
+    """原有参数名保持兼容。 / Preserve legacy bilinear parameter names."""
+    def __init__(self, hidden_size, projection_dim=256):
+        super().__init__()
+        self.decision = nn.Linear(hidden_size, projection_dim, bias=False)
+        self.candidate = nn.Linear(hidden_size, projection_dim, bias=False)
+        self.scale = math.sqrt(projection_dim)
+
+    def forward(self, hidden, positions, decision_position):
+        d = self.decision(hidden[0, decision_position].float())
+        c = self.candidate(hidden[0, positions].float())
+        return (c * d).sum(-1) / self.scale
+
+    def extract_features(self, hidden, branch):
+        return hidden[:, [*branch.candidate_positions, branch.decision_position], :]
+
+    def score_features(self, features):
+        return self(features, list(range(features.shape[1] - 1)), features.shape[1] - 1)
+
+
+class MLPHead(nn.Module):
+    """每个候选共享一个 MLP。 / One shared MLP for arbitrary candidate counts."""
+    def __init__(self, hidden_size, hidden_dim=512, bottleneck_dim=128):
+        super().__init__()
+        self.decision_norm = nn.LayerNorm(hidden_size)
+        self.candidate_norm = nn.LayerNorm(hidden_size)
+        self.network = nn.Sequential(nn.Linear(2 * hidden_size, hidden_dim), nn.GELU(),
+                                     nn.Linear(hidden_dim, bottleneck_dim), nn.GELU(),
+                                     nn.Linear(bottleneck_dim, 1))
+
+    def extract_features(self, hidden, branch):
+        c = hidden[0, branch.candidate_positions]
+        d = hidden[0, branch.decision_position].expand_as(c)
+        return torch.stack((d, c), dim=1)
+
+    def score_features(self, features):
+        d, c = features.float().unbind(1)
+        return self.network(torch.cat((self.decision_norm(d), self.candidate_norm(c)), -1)).squeeze(-1)
+
+
+def role_features(hidden, branch):
+    """先池化再投影，可缓存冻结特征。 / Pool before trainable projections."""
+    spans = (getattr(branch, "context_span", None), getattr(branch, "instruction_span", None))
+    candidates = getattr(branch, "candidate_spans", None)
+    if any(span is None for span in spans) or not candidates or len(candidates) != len(branch.candidate_positions):
+        raise ValueError("Role heads require compiler context, instruction and candidate spans")
+
+    def pool(span):
+        start, end = span
+        if not 0 <= start < end <= hidden.shape[1]:
+            raise ValueError(f"Invalid role span: {span}")
+        return hidden[0, start:end].float().mean(0)
+
+    context, instruction = (pool(span) for span in spans)
+    decision = hidden[0, branch.decision_position].float()
+    return torch.stack([torch.stack((context, instruction, pool(span), decision)) for span in candidates])
+
+
+class RoleHead(nn.Module):
+    def __init__(self, hidden_size, width):
+        super().__init__()
+        self.input_norm = nn.LayerNorm(hidden_size)
+        self.projection = nn.Linear(hidden_size, width)
+        self.roles = nn.Parameter(torch.zeros(1, 4, width))
+
+    def extract_features(self, hidden, branch):
+        return role_features(hidden, branch)
+
+    def project(self, features):
+        return self.projection(self.input_norm(features.float())) + self.roles
+
+
+class RoleMLPHead(RoleHead):
+    """与 Mixer 使用相同输入。 / Control for the Mixer's richer input features."""
+    def __init__(self, hidden_size, width=256, hidden_dim=1024):
+        super().__init__(hidden_size, width)
+        self.network = nn.Sequential(nn.Linear(4 * width, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, 1))
+
+    def score_features(self, features):
+        return self.network(self.project(features).flatten(1)).squeeze(-1)
+
+
+class MixerBlock(nn.Module):
+    def __init__(self, width, token_hidden, channel_hidden):
+        super().__init__()
+        self.token_norm = nn.LayerNorm(width)
+        self.token_mlp = nn.Sequential(nn.Linear(4, token_hidden), nn.GELU(), nn.Linear(token_hidden, 4))
+        self.channel_norm = nn.LayerNorm(width)
+        self.channel_mlp = nn.Sequential(nn.Linear(width, channel_hidden), nn.GELU(), nn.Linear(channel_hidden, width))
+
+    def forward(self, x):
+        # 混合角色，不混合不同候选。 / Mix fixed roles, never candidate rows.
+        x = x + self.token_mlp(self.token_norm(x).transpose(1, 2)).transpose(1, 2)
+        return x + self.channel_mlp(self.channel_norm(x))
+
+
+class MixerHead(RoleHead):
+    def __init__(self, hidden_size, width=256, layers=2, token_hidden=16, channel_hidden=1024):
+        super().__init__(hidden_size, width)
+        self.blocks = nn.Sequential(*(MixerBlock(width, token_hidden, channel_hidden) for _ in range(layers)))
+        self.output = nn.Sequential(nn.LayerNorm(width), nn.Linear(width, 1))
+
+    def score_features(self, features):
+        # 第三行是候选角色。 / Read the updated candidate role.
+        return self.output(self.blocks(self.project(features))[:, 2]).squeeze(-1)
+
+
+def build_head(hidden_size, config):
+    settings = head_signature(config)
+    kind = settings.pop("head_type")
+    if kind == "bilinear":
+        return DecisionHead(hidden_size, settings["projection_dim"])
+    if kind == "mlp":
+        return MLPHead(hidden_size, settings["head_hidden_dim"], settings["head_bottleneck_dim"])
+    if kind == "role_mlp":
+        return RoleMLPHead(hidden_size, settings["head_width"], settings["head_hidden_dim"])
+    return MixerHead(hidden_size, settings["head_width"], settings["head_layers"],
+                     settings["head_token_hidden_dim"], settings["head_channel_hidden_dim"])

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21fef5c5b93c2d9f1e54100f31709aaeb1651058784b9d38c47d9345fdba50d0
+size 3991299376

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ec6792a776db86b348b97ae3e9ff5feaa055dba2ea28035aac9f1d543623e9b
+size 3979833864

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7b96a0d85eb01475c0f775a51f88e4455aeae31143cc86c8457723d9019c13e
+size 1114356012

model.safetensors.index.json

@@ -0,0 +1,764 @@
+{
+  "metadata": {
+    "total_parameters": 4540981545,
+    "total_size": 9085395108
+  },
+  "weight_map": {
+    "backbone.language_model.embed_tokens.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.0.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.1.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.10.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.self_attn.k_norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.self_attn.q_norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.11.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.12.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.12.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.12.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.12.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.12.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.13.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.14.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.self_attn.k_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.self_attn.q_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.15.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.16.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.17.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.18.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.self_attn.k_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.self_attn.q_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.19.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.2.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.2.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.20.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.20.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.21.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.22.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.self_attn.k_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.self_attn.q_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.23.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.24.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.25.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.26.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.self_attn.k_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.self_attn.q_norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.27.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.28.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.in_proj_qkv.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.in_proj_z.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.norm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.linear_attn.out_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.29.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.3.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.self_attn.k_norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.self_attn.q_norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.3.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.30.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.A_log": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.conv1d.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.dt_bias": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.in_proj_a.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.in_proj_b.weight": "model-00002-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.in_proj_qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.in_proj_z.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.norm.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.linear_attn.out_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.30.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.self_attn.k_norm.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.self_attn.q_norm.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.31.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.language_model.layers.4.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.4.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.5.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.6.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.self_attn.k_norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.self_attn.q_norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.7.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.8.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.A_log": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.conv1d.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.dt_bias": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.in_proj_a.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.in_proj_b.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.in_proj_qkv.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.in_proj_z.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.norm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.linear_attn.out_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.layers.9.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "backbone.language_model.norm.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.0.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.1.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.10.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.11.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.12.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.13.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.14.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.15.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.16.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.17.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.18.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.19.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.2.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.20.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.21.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.22.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.23.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.3.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.4.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.5.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.6.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.7.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.8.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.attn.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.attn.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.attn.qkv.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.attn.qkv.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.mlp.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.mlp.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.mlp.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.mlp.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.norm1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.norm1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.norm2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.blocks.9.norm2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.merger.linear_fc1.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.merger.linear_fc1.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.merger.linear_fc2.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.merger.linear_fc2.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.merger.norm.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.merger.norm.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.patch_embed.proj.bias": "model-00003-of-00003.safetensors",
+    "backbone.visual.patch_embed.proj.weight": "model-00003-of-00003.safetensors",
+    "backbone.visual.pos_embed.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.0.channel_mlp.0.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.0.channel_mlp.0.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.0.channel_mlp.2.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.0.channel_mlp.2.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.0.channel_norm.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.0.channel_norm.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.0.token_mlp.0.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.0.token_mlp.0.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.0.token_mlp.2.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.0.token_mlp.2.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.0.token_norm.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.0.token_norm.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.1.channel_mlp.0.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.1.channel_mlp.0.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.1.channel_mlp.2.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.1.channel_mlp.2.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.1.channel_norm.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.1.channel_norm.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.1.token_mlp.0.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.1.token_mlp.0.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.1.token_mlp.2.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.1.token_mlp.2.weight": "model-00003-of-00003.safetensors",
+    "head.blocks.1.token_norm.bias": "model-00003-of-00003.safetensors",
+    "head.blocks.1.token_norm.weight": "model-00003-of-00003.safetensors",
+    "head.input_norm.bias": "model-00003-of-00003.safetensors",
+    "head.input_norm.weight": "model-00003-of-00003.safetensors",
+    "head.output.0.bias": "model-00003-of-00003.safetensors",
+    "head.output.0.weight": "model-00003-of-00003.safetensors",
+    "head.output.1.bias": "model-00003-of-00003.safetensors",
+    "head.output.1.weight": "model-00003-of-00003.safetensors",
+    "head.projection.bias": "model-00003-of-00003.safetensors",
+    "head.projection.weight": "model-00003-of-00003.safetensors",
+    "head.roles": "model-00003-of-00003.safetensors"
+  }
+}

modeling_valen.py

@@ -0,0 +1,84 @@
+"""Complete merged Qwen + decision head. / 已合并的完整 Qwen 和决策头。"""
+import torch
+from transformers import AutoProcessor
+from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5Model, Qwen3_5PreTrainedModel
+
+from .configuration_valen import ValenQwenConfig
+from .heads import build_head
+from .runtime_model import ValenQwen
+from .compiler import Compiler
+from .responses import answer
+# List transitive files explicitly for Transformers' local dynamic loader.
+# 本地动态加载需要显式列出间接依赖，否则搬移后可能缺少这些模块。
+from .batching import collate_branches
+from .schema import candidates
+from .data_types import QuestionSpec
+
+
+class ValenQwenForDecisionMaking(Qwen3_5PreTrainedModel):
+    config_class = ValenQwenConfig
+    base_model_prefix = "backbone"
+    model_name = "Valen"
+    architecture = "qwen"
+    # The trained Mixer stays FP32 with a BF16 backbone. / BF16 基座、FP32 决策头。
+    _keep_in_fp32_modules_strict = ["head"]
+
+    def __init__(self, config, backbone=None, head=None):
+        super().__init__(config)
+        if (backbone is None) != (head is None):
+            raise ValueError("Supply both a merged backbone and a trained head")
+        self.backbone = backbone if backbone is not None else Qwen3_5Model(config)
+        self.head = head if head is not None else build_head(config.text_config.hidden_size, config.valen_head)
+        self.head.float()
+        self._processor = None
+        # Supplied weights are already trained: do not reinitialize them.
+        # 传入的参数已训练完成，不能再次调用初始化覆盖它们。
+        if backbone is not None:
+            for module in self.modules():
+                module._is_hf_initialized = True
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.backbone.set_input_embeddings(value)
+
+    def forward(self, question=None, **inputs):
+        if question is None:
+            return self.backbone(**inputs)
+        if hasattr(question, "questions"):
+            return self.forward_state(question)
+        return ValenQwen.forward(self, question)
+
+    # Use the exact trained readouts and batching implementation.
+    # 保留训练时的特征读取和 batch 逻辑。
+    extract_features = ValenQwen.extract_features
+    extract_state_features = ValenQwen.extract_state_features
+    _state_features = ValenQwen._state_features
+    forward_state = ValenQwen.forward_state
+    forward_state_batch = ValenQwen.forward_state_batch
+    forward_batch = ValenQwen.forward_batch
+    score_features = ValenQwen.score_features
+
+    def make_compiler(self, media_root=".", execution=None, processor=None):
+        if processor is None:
+            if self._processor is None:
+                self._processor = AutoProcessor.from_pretrained(self.config._name_or_path, trust_remote_code=True)
+            processor = self._processor
+        return Compiler(processor, media_root, self.config.valen_max_length, self.config.valen_media_kwargs,
+                        execution=execution or self.config.qwen_execution)
+
+    @torch.inference_mode()
+    def predict(self, request, *, media_root=".", execution=None, temperature=1.0, processor=None):
+        self.eval()
+        # Only request content reaches the compiler. / 只编译输入，不读取标签。
+        request = request["request"] if "request" in request else request
+        compiler = self.make_compiler(media_root, execution, processor)
+        state = compiler.compile({"request": request})
+        logits = self.forward_state(state) if state.inputs is not None else [self(q) for q in state.questions]
+        answers = {q.qid: answer(q, z, temperature[q.kind] if isinstance(temperature, dict) else temperature)
+                   for q, z in zip(state.questions, logits)}
+        return {"model": self.model_name, "answers": answers,
+                "usage": {"input_tokens": state.logical_tokens, "output_tokens": 0},
+                "internal_usage": {"compute_tokens": state.compute_tokens}}

processor_config.json

@@ -0,0 +1,60 @@
+{
+  "image_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "merge_size": 2,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "size": {
+      "longest_edge": 16777216,
+      "shortest_edge": 65536
+    },
+    "temporal_patch_size": 2
+  },
+  "processor_class": "Qwen3VLProcessor",
+  "video_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "do_sample_frames": true,
+    "fps": 2,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_frames": 768,
+    "merge_size": 2,
+    "min_frames": 4,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "return_metadata": false,
+    "size": {
+      "longest_edge": 25165824,
+      "shortest_edge": 4096
+    },
+    "temporal_patch_size": 2,
+    "video_processor_type": "Qwen3VLVideoProcessor"
+  }
+}

responses.py

@@ -0,0 +1,29 @@
+import math
+import torch
+
+def answer(question, logits, temperature=1.0):
+    if not math.isfinite(temperature) or temperature <= 0:
+        raise ValueError("Calibration temperature must be finite and positive")
+    if logits.ndim != 1 or len(logits) != len(question.keys) or not torch.isfinite(logits).all():
+        raise ValueError("Expected one finite logit per candidate")
+    probs = (logits.float() / temperature).softmax(-1).detach().cpu().tolist()
+    total = sum(probs)
+    probs = [p / total for p in probs]
+    result = {"type": question.kind}
+    if question.kind == "noul":
+        result["noul"] = probs[question.keys.index("true")]
+        return result
+    result["probabilities"] = dict(zip(question.keys, probs))
+    mode = max(range(len(probs)), key=probs.__getitem__)
+    if question.kind == "choice":
+        result["choice"] = question.keys[mode]
+        result["confidence"] = 1.0 if len(probs) == 1 else max(0.0, (max(probs) - 1/len(probs)) / (1 - 1/len(probs)))
+    else:
+        result["score"] = sum(i * p for i, p in enumerate(probs))
+        result["legend"] = dict(zip(question.keys, question.descriptions))
+        # Independently implemented formula from pinned system-one adapter:
+        # fb52b1030b7fc1f4f1cf39910afa5da54f9835e3, confidence_metrics.py.
+        distance = sum(p * abs(i - mode) for i, p in enumerate(probs))
+        uniform_mad = sum(abs(i - (len(probs)-1)/2) for i in range(len(probs))) / len(probs)
+        result["confidence"] = max(0.0, 1.0 - distance / uniform_mad)
+    return result

runtime_model.py

@@ -0,0 +1,101 @@
+import torch
+from torch import nn
+MODEL_NAME = "Valen"
+from .heads import DecisionHead, build_head
+from .batching import branch_batches, collate_branches
+
+
+class ValenQwen(nn.Module):
+    """Qwen3.5 backbone with a shared candidate decision head."""
+
+    model_name = MODEL_NAME
+    architecture = "qwen"
+
+    def __init__(self, backbone, projection_dim=256, head_config=None):
+        super().__init__()
+        self.backbone = backbone
+        self.head = build_head(backbone.config.text_config.hidden_size,
+                               dict({"projection_dim": projection_dim}, **(head_config or {})))
+
+    def forward(self, question):
+        device = next(self.backbone.parameters()).device
+        outputs = []
+        for branch in question.branches:
+            inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in branch.inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            outputs.append(self.head.score_features(self.head.extract_features(hidden, branch)))
+        return torch.cat(outputs)
+
+    def extract_features(self, question):
+        """缓存投影前的特征。 / Cache raw readouts before trainable head layers."""
+        device = next(self.backbone.parameters()).device
+        features = []
+        for branch in question.branches:
+            inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in branch.inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            features.append(self.head.extract_features(hidden, branch))
+        return features
+
+    def extract_state_features(self, state):
+        """一次编码 state，读取各题特征。 / One backbone graph for every question."""
+        if not state.questions:
+            return []
+        if state.inputs is None:
+            raise ValueError("Shared-state features require compiled shared inputs")
+        device = next(self.backbone.parameters()).device
+        inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in state.inputs.items()}
+        hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+        return self._state_features(hidden, state)
+
+    def _state_features(self, hidden, state):
+        return [[self.head.extract_features(hidden, branch) for branch in q.branches]
+                for q in state.questions]
+
+    def forward_state(self, state):
+        return [self.score_features(features) for features in self.extract_state_features(state)]
+
+    def forward_state_batch(self, states, max_tokens=32768):
+        """不同 state 组成 batch，同一 state 的题目共用一行。 / One row per shared state."""
+        device = next(self.backbone.parameters()).device
+        pad_token_id = getattr(self.backbone.config.text_config, "pad_token_id", None) or 0
+        outputs = [[] for _ in states]
+        representatives = []
+        indices = []
+        for index, state in enumerate(states):
+            if not state.questions:
+                continue
+            if state.inputs is None or any(branch.inputs is not state.inputs
+                    for q in state.questions for branch in q.branches):
+                raise ValueError("Shared-state batch requires one shared input per state")
+            representatives.append(state.questions[0])
+            indices.append(index)
+        for batch in branch_batches(representatives, max_tokens):
+            inputs = collate_branches([branch for _, branch in batch], pad_token_id)
+            inputs = {key: value.to(device) for key, value in inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            for row, (index, _) in enumerate(batch):
+                original = indices[index]
+                features = self._state_features(hidden[row:row + 1], states[original])
+                outputs[original] = [self.score_features(f) for f in features]
+        return outputs
+
+    def forward_batch(self, questions, max_tokens=32768):
+        """Forward several QAs together; each keeps its own decision head inputs.
+
+        多条 QA 并行；每条保留独立的候选、角色区间和 Score 分支。
+        """
+        device = next(self.backbone.parameters()).device
+        pad_token_id = getattr(self.backbone.config.text_config, "pad_token_id", None) or 0
+        outputs = [[] for _ in questions]
+        for batch in branch_batches(questions, max_tokens):
+            inputs = collate_branches([branch for _, branch in batch], pad_token_id)
+            inputs = {key: value.to(device) for key, value in inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            for row, (question_index, branch) in enumerate(batch):
+                features = self.head.extract_features(hidden[row:row + 1], branch)
+                outputs[question_index].append(self.head.score_features(features))
+        return [torch.cat(branches) for branches in outputs]
+
+    def score_features(self, features, head=None):
+        head = self.head if head is None else head
+        return torch.cat([head.score_features(hidden) for hidden in features])

schema.py

@@ -0,0 +1,71 @@
+"""Strict JSONL contract. IDs and labels never enter the model prompt."""
+import json
+import math
+from pathlib import Path
+
+
+def candidates(question):
+    kind = question.get("type")
+    if not isinstance(question.get("instructions"), str) or not question["instructions"].strip():
+        raise ValueError("instructions must be nonempty text")
+    criteria = question.get("criteria")
+    if kind == "noul":
+        return [("true", "True / 是：满足问题中的条件。"), ("false", "False / 否：不满足问题中的条件。")]
+    if kind == "choice":
+        if not isinstance(criteria, dict) or not 1 <= len(criteria) <= 255:
+            raise ValueError("Choice needs 1–255 named criteria")
+        result = list(criteria.items())
+    elif kind == "score":
+        if not isinstance(criteria, list) or not 2 <= len(criteria) <= 255:
+            raise ValueError("Score criteria must be an ordered list of 2–255 descriptions")
+        result = [(str(i), text) for i, text in enumerate(criteria)]
+    else:
+        raise ValueError(f"Unsupported question type: {kind}")
+    if any(not isinstance(k, str) or not k or not isinstance(v, str) or not v.strip() for k, v in result):
+        raise ValueError("Candidate names and descriptions must be nonempty strings")
+    return result
+
+
+def target_distribution(target, keys):
+    if target is None or target.get("probabilities") is None:
+        return None
+    probs = target["probabilities"]
+    if not isinstance(probs, dict) or set(probs) != set(keys):
+        raise ValueError(f"Probability keys must exactly match {keys}")
+    values = [probs[k] for k in keys]
+    if any(isinstance(v, bool) or not isinstance(v, (int, float)) or not math.isfinite(v) or v < 0 for v in values):
+        raise ValueError("Probabilities must be finite nonnegative numbers")
+    if not math.isclose(sum(values), 1.0, abs_tol=1e-6):
+        raise ValueError("Target probabilities must sum to 1")
+    return values
+
+
+def validate_record(record, candidate_fn=candidates):
+    request = record["request"]
+    state = request["state"]
+    if not isinstance(state, (str, dict)):
+        raise ValueError("state must be text or a messages object")
+    if not isinstance(request["questions"], dict) or not request["questions"]:
+        raise ValueError("questions must be a nonempty mapping")
+    targets = record.get("targets", {})
+    if set(targets) - set(request["questions"]):
+        raise ValueError("Unknown target question IDs")
+    for qid, question in request["questions"].items():
+        target_distribution(targets.get(qid), [k for k, _ in candidate_fn(question)])
+    if not record.get("group_id"):
+        raise ValueError("group_id is required for leakage-free splits")
+    return record
+
+
+def read_jsonl(path, candidate_fn=candidates):
+    records = []
+    with Path(path).open(encoding="utf-8") as stream:
+        for line_number, line in enumerate(stream, 1):
+            if line.strip():
+                try:
+                    records.append(validate_record(json.loads(line), candidate_fn))
+                except (ValueError, KeyError, TypeError, AttributeError) as exc:
+                    raise ValueError(f"{path}:{line_number}: {exc}") from exc
+    if not records:
+        raise ValueError("Empty dataset")
+    return records

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4
+size 19989343

tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "processor_class": "Qwen3VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

unmerged/COMPLETE.json

@@ -0,0 +1,4 @@
+{
+  "validated": true,
+  "source_checkpoint_sha256": "b4193b4e907f44b913fd54a8d6b998e7050a1528306269a000f1e5e8bd8dffc7"
+}

unmerged/LICENSE

@@ -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 2026 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.

unmerged/README.md

@@ -0,0 +1,50 @@
+# Valen-4B — unmerged LoRA
+
+原始 Qwen3.5 基座 + 未合并的 LoRA + FP32 Mixer + 训练后的视觉 merger。
+Original base with separate LoRA, Mixer, and trained visual merger weights.
+来源：1195k + JevBench 30k，两阶段 SFT，最终 step 325。
+
+## 文件 / Files
+
+- `lora/adapter_model.safetensors`、`lora/adapter_config.json`：标准 PEFT LoRA。
+- `mixer.safetensors`：完整 FP32 Mixer 决策头。
+- `visual_merger.safetensors`：联合微调时训练的视觉 merger，加载时一并恢复。
+- `config.json`、processor、tokenizer 和 Python 文件：模型配置及独立加载代码。
+
+## 完整模型加载 / Load the complete decision model
+
+环境沿用父目录：PyTorch 2.6.0、Transformers 5.4.0，另需 PEFT 0.18.1 和 safetensors。
+Flash Attention 2 可通过 `attn_implementation="flash_attention_2"` 启用。
+
+```python
+import torch
+from transformers import AutoModel
+
+model = AutoModel.from_pretrained(
+    "./Valen-4B/unmerged",
+    base_model_path="./Qwen3.5-4B",  # 原始官方基座 / original published base
+    trust_remote_code=True,
+    dtype=torch.bfloat16,
+    attn_implementation="sdpa",
+).to("cuda").eval()
+torch.set_float32_matmul_precision("highest")
+torch.backends.cudnn.allow_tf32 = False
+
+result = model.predict({
+    "state": "There is one cat in the room.",
+    "questions": {
+        "animal": {"type": "choice", "instructions": "Which animal is present?",
+                   "criteria": {"cat": "A cat", "dog": "A dog"}},
+        "is_cat": {"type": "noul", "instructions": "There is a cat in the room."},
+    },
+})
+print(result["answers"])
+```
+
+`base_model_path` 指向原始官方 `Qwen/Qwen3.5-4B` 权重，本目录中的加载器会核对基座 SHA-256。
+省略该参数时使用配置记录的仓库及固定 revision；离线环境需提前缓存该版本。
+默认 share_state，也支持 `model.predict(..., execution="question")`；视频采样 16 帧。
+内部使用 `PeftModel.from_pretrained(backbone.language_model, .../lora)` 加载语言 LoRA，随后恢复 Mixer 和视觉 merger。
+适配器不会被融合。`is_trainable=True` 可启用 LoRA、Mixer 和视觉 merger 的梯度。
+
+`validation.json` 记录恢复后与原 checkpoint 的一致性；文件来源和校验和见 `export_manifest.json`。

unmerged/batching.py

@@ -0,0 +1,62 @@
+"""Pad language rows and concatenate visual patches. / 文本补齐，视觉 patch 拼接。"""
+import torch
+from torch.nn.utils.rnn import pad_sequence
+
+
+SEQUENCE_KEYS = {"input_ids", "attention_mask", "mm_token_type_ids"}
+VISUAL_KEYS = {"pixel_values", "pixel_values_videos", "image_grid_thw", "video_grid_thw"}
+
+
+def collate_branches(branches, pad_token_id=0):
+    if not branches:
+        raise ValueError("Cannot collate an empty branch batch")
+    inputs = [branch.inputs for branch in branches]
+    unknown = set().union(*(row.keys() for row in inputs)) - SEQUENCE_KEYS - VISUAL_KEYS
+    if unknown:
+        raise ValueError(f"Unsupported batched Qwen inputs: {sorted(unknown)}")
+    if any(row["input_ids"].ndim != 2 or row["input_ids"].shape[0] != 1 for row in inputs):
+        raise ValueError("Each compiled branch must contain one language row")
+    result = {}
+    for key in SEQUENCE_KEYS:
+        if key != "input_ids" and not any(key in row for row in inputs):
+            continue
+        rows = []
+        for row in inputs:
+            value = row.get(key)
+            if value is None:
+                value = (torch.ones_like(row["input_ids"]) if key == "attention_mask"
+                         else torch.zeros_like(row["input_ids"]))
+            if value.shape != row["input_ids"].shape:
+                raise ValueError(f"{key} must match the branch input_ids shape")
+            rows.append(value[0])
+        # Right padding keeps all compiler readout positions and spans unchanged.
+        # 右侧补齐，候选位置和角色区间无需偏移。
+        result[key] = pad_sequence(rows, batch_first=True,
+                                   padding_value=pad_token_id if key == "input_ids" else 0)
+    if "attention_mask" not in result:
+        result["attention_mask"] = pad_sequence([torch.ones_like(r["input_ids"][0]) for r in inputs],
+                                                batch_first=True, padding_value=0)
+    for key in VISUAL_KEYS:
+        values = [row[key] for row in inputs if key in row]
+        if values:
+            result[key] = torch.cat(values, dim=0)
+    return result
+
+
+def branch_batches(questions, max_tokens):
+    """Limit padded tokens per backbone forward, keeping Score branch order."""
+    if type(max_tokens) is not int or max_tokens <= 0:
+        raise ValueError("microbatch_max_tokens must be a positive integer")
+    batch, longest = [], 0
+    for question_index, question in enumerate(questions):
+        if not question.branches:
+            raise ValueError("A Qwen question requires at least one branch")
+        for branch in question.branches:
+            length = branch.inputs["input_ids"].shape[1]
+            if batch and max(longest, length) * (len(batch) + 1) > max_tokens:
+                yield batch
+                batch, longest = [], 0
+            batch.append((question_index, branch))
+            longest = max(longest, length)
+    if batch:
+        yield batch

unmerged/chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

unmerged/compiler.py

@@ -0,0 +1,239 @@
+"""Compile Qwen question branches or a shared sequence with per-question readouts."""
+from copy import deepcopy
+from dataclasses import dataclass
+from pathlib import Path
+import hashlib
+import torch
+from .schema import candidates, target_distribution
+from .data_types import QuestionSpec
+
+
+@dataclass
+class Branch:
+    inputs: dict
+    candidate_positions: list
+    decision_position: int
+    context_span: tuple | None = None
+    instruction_span: tuple | None = None
+    candidate_spans: list | None = None
+
+
+@dataclass(init=False)
+class Question(QuestionSpec):
+    branches: list
+
+    def __init__(self, qid, kind, keys, descriptions, branches, target, is_text):
+        # Keep the positional constructor used by existing compiler clients.
+        super().__init__(qid, kind, keys, descriptions, target, is_text)
+        self.branches = branches
+
+
+@dataclass
+class CompiledState:
+    questions: list
+    logical_tokens: int
+    compute_tokens: int
+    media: list
+    inputs: dict | None = None
+
+
+class Compiler:
+    def __init__(self, processor, media_root=".", max_length=8192, media_kwargs=None, verified_media_hashes=None,
+                 execution="question"):
+        if execution not in {"question", "shared_state"}:
+            raise ValueError(f"Unknown qwen_execution: {execution}")
+        self.execution = execution
+        self.processor = processor
+        self.tokenizer = processor.tokenizer
+        self.media_root = Path(media_root)
+        self.max_length = max_length
+        self.media_kwargs = media_kwargs or {}
+        # Optional, pre-audited immutable assets for latency measurements.
+        # Normal training/inference continues to hash each input at use time.
+        self.verified_media_hashes = verified_media_hashes or {}
+
+    def _messages(self, state):
+        messages = [{"role": "user", "content": state}] if isinstance(state, str) else deepcopy(state["messages"])
+        media = []
+        for message in messages:
+            if message["role"] not in {"user", "assistant", "system"}:
+                raise ValueError("Unsupported state role")
+            if isinstance(message["content"], str):
+                message["content"] = [{"type": "text", "text": message["content"]}]
+            for item in message["content"]:
+                kind = item["type"]
+                if kind == "text":
+                    # Prevent user text from injecting Qwen control tokens.
+                    if any(t in item["text"] for t in self.tokenizer.all_special_tokens):
+                        raise ValueError("Reserved tokenizer control token in state")
+                    continue
+                if kind not in {"image_url", "video_url"}:
+                    raise ValueError(f"Unsupported content type: {kind}")
+                url = item[kind]["url"]
+                if "://" in url:
+                    raise ValueError("Materialize remote media locally with a hash before training")
+                path = Path(url)
+                path = path if path.is_absolute() else self.media_root / path
+                path = path.resolve(strict=True)
+                digest = self.verified_media_hashes.get(str(path))
+                if digest is None:
+                    digest = hashlib.sha256(path.read_bytes()).hexdigest()
+                media.append({"path": str(path), "sha256": digest, "type": kind})
+                item.clear()
+                item.update(type=kind.removesuffix("_url"), path=str(path))
+        return messages, media
+
+    def compile(self, record, rng=None, labeled_only=False):
+        selected = []
+        for qid, question in record["request"]["questions"].items():
+            pairs = candidates(question)
+            target = record.get("targets", {}).get(qid)
+            if labeled_only and self.execution == "question" and target_distribution(target, [k for k, _ in pairs]) is None:
+                continue
+            if rng is not None and question["type"] == "choice":
+                rng.shuffle(pairs)
+            selected.append((qid, question, pairs, target))
+        # 标签只决定监督范围，共享序列仍保留所有题。 / Labels select losses, never prompt content.
+        if not selected or (labeled_only and not any(
+                target_distribution(target, [k for k, _ in pairs]) is not None
+                for _, _, pairs, target in selected)):
+            return CompiledState([], 0, 0, [])
+        messages, media = self._messages(record["request"]["state"])
+        expected_assets = {str((self.media_root / a["path"]).resolve()): a["sha256"] for a in record.get("assets", [])}
+        for entry in media:
+            if entry["path"] in expected_assets and expected_assets[entry["path"]] != entry["sha256"]:
+                raise ValueError(f"Media hash mismatch: {entry['path']}")
+        base = dict(self.processor.apply_chat_template(
+            messages, tokenize=True, add_generation_prompt=False, return_dict=True,
+            return_tensors="pt", processor_kwargs=dict(self.media_kwargs, return_metadata=True)))
+        # Metadata is audit information, never a backbone keyword argument.
+        metadata = base.pop("video_metadata", None)
+        if metadata is not None:
+            for entry, meta in zip([m for m in media if m["type"] == "video_url"], metadata):
+                entry["sampling"] = dict(meta) if isinstance(meta, dict) else vars(meta)
+                sampling = entry["sampling"]
+                indices = sampling.get("frames_indices")
+                if indices is not None and sampling.get("fps"):
+                    sampling["frames_indices"] = [int(i) for i in indices]
+                    sampling["timestamps_seconds"] = [int(i) / sampling["fps"] for i in indices]
+                    sampling["visible_range_seconds"] = [min(sampling["timestamps_seconds"]), max(sampling["timestamps_seconds"])]
+        for name in ("image_grid_thw", "video_grid_thw"):
+            if name in base:
+                grid = base[name]
+                media.append({name: grid.tolist(), "visual_tokens": int(grid.prod(-1).sum()) // 4})
+        base_length = base["input_ids"].shape[1]
+        if self.execution == "shared_state":
+            return self._compile_shared(selected, base, media, labeled_only)
+        logical_tokens, compute_tokens = base_length, 0
+        compiled = []
+        for qid, question, pairs, target in selected:
+            keys = [k for k, _ in pairs]
+            branches = []
+            groups = [[pair] for pair in pairs] if question["type"] == "score" else [pairs]
+            for group in groups:
+                suffix = []
+                def append(text, char_span=None):
+                    start = base_length + len(suffix)
+                    if char_span is None:
+                        suffix.extend(self.tokenizer.encode(text, add_special_tokens=False))
+                        return None
+                    # 保持原有分词，只额外记录区间。 / Preserve token IDs; record role spans.
+                    encoded = self.tokenizer(text, add_special_tokens=False, return_offsets_mapping=True)
+                    suffix.extend(encoded["input_ids"])
+                    indices = [i for i, (a, b) in enumerate(encoded["offset_mapping"])
+                               if a < char_span[1] and b > char_span[0]]
+                    if not indices:
+                        raise ValueError("Empty token span for decision role")
+                    return (start + indices[0], start + indices[-1] + 1)
+                # Canonical segment tokenization makes endpoints exact; no string searching.
+                suffix.extend(self.tokenizer.encode("<|im_start|>user\n", add_special_tokens=False))
+                for value in [question["instructions"]] + [v for pair in group for v in pair]:
+                    if any(t in value for t in self.tokenizer.all_special_tokens):
+                        raise ValueError("Reserved tokenizer control token in question/criteria")
+                prefix = "Task: " + question["type"] + "\nQuestion: "
+                instruction_span = append(prefix + question["instructions"] + "\nCandidates:\n",
+                                          (len(prefix), len(prefix) + len(question["instructions"])))
+                positions, candidate_spans = [], []
+                for key, description in group:
+                    # Score never sees its level ID, nor any other level description.
+                    prefix = key + ": " if question["type"] != "score" else ""
+                    candidate_spans.append(append(prefix + description, (len(prefix), len(prefix) + len(description))))
+                    positions.append(base_length + len(suffix) - 1)
+                    append("\n")
+                append("Decision:")
+                length = base_length + len(suffix)
+                if length > self.max_length:
+                    raise ValueError(f"{qid}: {length} tokens exceed max_length={self.max_length}; no truncation")
+                inputs = dict(base)
+                inputs["input_ids"] = torch.cat([base["input_ids"], torch.tensor([suffix], dtype=torch.long)], dim=1)
+                inputs["attention_mask"] = torch.ones_like(inputs["input_ids"])
+                if "mm_token_type_ids" in base:
+                    inputs["mm_token_type_ids"] = torch.cat([base["mm_token_type_ids"], torch.zeros((1, len(suffix)), dtype=torch.long)], dim=1)
+                branches.append(Branch(inputs, positions, length - 1, (0, base_length),
+                                       instruction_span, candidate_spans))
+                compute_tokens += length
+                logical_tokens += len(suffix)
+            compiled.append(Question(qid, question["type"], keys, [v for _, v in pairs], branches,
+                                     target_distribution(target, keys), not media))
+        return CompiledState(compiled, logical_tokens, compute_tokens, media)
+
+    def _compile_shared(self, selected, base, media, labeled_only=False):
+        """All schemas precede all readouts; labels and external IDs never enter inputs.
+
+        This is ordinary causal attention, not isolated question/Score branches.
+        No answer tokens or LM output layer are required by the decision heads.
+        """
+        base_length = base["input_ids"].shape[1]
+        suffix, compiled = [], []
+
+        def append(text, span=None):
+            start = base_length + len(suffix)
+            if span is None:
+                suffix.extend(self.tokenizer.encode(text, add_special_tokens=False))
+                return None
+            encoded = self.tokenizer(text, add_special_tokens=False, return_offsets_mapping=True)
+            suffix.extend(encoded["input_ids"])
+            indices = [i for i, (a, b) in enumerate(encoded["offset_mapping"]) if a < span[1] and b > span[0]]
+            if not indices:
+                raise ValueError("Empty token span for decision role")
+            return start + indices[0], start + indices[-1] + 1
+
+        append("<|im_start|>user\n")
+        for index, (qid, question, pairs, target) in enumerate(selected):
+            for value in [question["instructions"]] + [v for pair in pairs for v in pair]:
+                if any(t in value for t in self.tokenizer.all_special_tokens):
+                    raise ValueError("Reserved tokenizer control token in question/criteria")
+            prefix = f"Question {index + 1}\nTask: {question['type']}\nQuestion: "
+            instruction = append(prefix + question["instructions"] + "\nCandidates:\n",
+                                 (len(prefix), len(prefix) + len(question["instructions"])))
+            positions, spans = [], []
+            for key, description in pairs:
+                prefix = key + ": " if question["type"] != "score" else ""
+                spans.append(append(prefix + description, (len(prefix), len(prefix) + len(description))))
+                positions.append(base_length + len(suffix) - 1)
+                append("\n")
+            # Inputs and the decision position are filled after all schemas are appended.
+            branch = Branch({}, positions, -1, (0, base_length), instruction, spans)
+            keys = [key for key, _ in pairs]
+            compiled.append(Question(qid, question["type"], keys, [v for _, v in pairs], [branch],
+                                     target_distribution(target, keys), not media))
+        append("<|im_end|>\n<|im_start|>assistant\n")
+        for index, question in enumerate(compiled):
+            append(f"Question {index + 1} Decision:")
+            question.branches[0].decision_position = base_length + len(suffix) - 1
+            append("\n")
+        length = base_length + len(suffix)
+        if length > self.max_length:
+            raise ValueError(f"Shared state: {length} tokens exceed max_length={self.max_length}; no truncation")
+        inputs = dict(base)
+        inputs["input_ids"] = torch.cat([base["input_ids"], base["input_ids"].new_tensor([suffix])], dim=1)
+        inputs["attention_mask"] = torch.cat([base["attention_mask"],
+                                             base["attention_mask"].new_ones((1, len(suffix)))], dim=1)
+        if "mm_token_type_ids" in base:
+            inputs["mm_token_type_ids"] = torch.cat([base["mm_token_type_ids"],
+                                                    base["mm_token_type_ids"].new_zeros((1, len(suffix)))], dim=1)
+        for question in compiled:
+            question.branches[0].inputs = inputs
+        if labeled_only:
+            compiled = [question for question in compiled if question.target is not None]
+        return CompiledState(compiled, length, length, media, inputs)

unmerged/config.json

@@ -0,0 +1,163 @@
+{
+  "architectures": [
+    "ValenQwenWithLoRA"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_valen.ValenQwenConfig",
+    "AutoModel": "modeling_valen_lora.ValenQwenWithLoRA"
+  },
+  "dtype": "bfloat16",
+  "image_token_id": 248056,
+  "model_type": "valen_qwen",
+  "qwen_execution": "shared_state",
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attn_output_gate": true,
+    "bos_token_id": null,
+    "dtype": "bfloat16",
+    "eos_token_id": 248044,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 2560,
+    "initializer_range": 0.02,
+    "intermediate_size": 9216,
+    "layer_types": [
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention"
+    ],
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 32,
+    "linear_value_head_dim": 128,
+    "mamba_ssm_dtype": "float32",
+    "max_position_embeddings": 262144,
+    "mlp_only_layers": [],
+    "model_type": "qwen3_5_text",
+    "mtp_num_hidden_layers": 1,
+    "mtp_use_dedicated_embeddings": false,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 4,
+    "pad_token_id": null,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        11,
+        11,
+        10
+      ],
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 10000000,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 248320
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.4.0",
+  "valen_head": {
+    "head_channel_hidden_dim": 1024,
+    "head_layers": 2,
+    "head_token_hidden_dim": 16,
+    "head_type": "mixer",
+    "head_width": 256
+  },
+  "valen_max_length": 32768,
+  "valen_media_kwargs": {
+    "images_kwargs": {
+      "size": {
+        "longest_edge": 262144,
+        "shortest_edge": 65536
+      }
+    },
+    "videos_kwargs": {
+      "fps": null,
+      "num_frames": 16,
+      "size": {
+        "longest_edge": 2097152,
+        "shortest_edge": 131072
+      }
+    }
+  },
+  "video_token_id": 248057,
+  "vision_config": {
+    "deepstack_visual_indexes": [],
+    "depth": 24,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1024,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4096,
+    "model_type": "qwen3_5",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 2560,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 248054,
+  "vision_start_token_id": 248053,
+  "valen_base": {
+    "repo": "Qwen/Qwen3.5-4B",
+    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+    "weight_sha256": {
+      "model.safetensors-00001-of-00002.safetensors": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
+      "model.safetensors-00002-of-00002.safetensors": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188"
+    },
+    "files_sha256": {
+      ".gitattributes": "34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930",
+      "LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
+      "README.md": "1406be1b6b8fd8a6545870da516912804756593628a1d0fb0a7965211e82a7bb",
+      "chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+      "config.json": "ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670",
+      "merges.txt": "a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d",
+      "model.safetensors-00001-of-00002.safetensors": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
+      "model.safetensors-00002-of-00002.safetensors": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188",
+      "model.safetensors.index.json": "cf3f798ee02ba45f9622aa8892a47369ab667d0afbf154ee7c2212de42e6302d",
+      "preprocessor_config.json": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
+      "tokenizer.json": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42",
+      "tokenizer_config.json": "316230d6a809701f4db5ea8f8fc862bc3a6f3229c937c174e674ff3ca0a64ac8",
+      "video_preprocessor_config.json": "7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13",
+      "vocab.json": "ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003"
+    },
+    "verified_from": "download_manifest.json; every LFS SHA256 matched"
+  }
+}

unmerged/configuration_valen.py

@@ -0,0 +1,16 @@
+"""Standalone Qwen decision configuration. / 独立的 Qwen 决策模型配置。"""
+from transformers import Qwen3_5Config
+
+
+class ValenQwenConfig(Qwen3_5Config):
+    model_type = "valen_qwen"
+
+    def __init__(self, valen_head=None, qwen_execution="shared_state", valen_max_length=32768,
+                 valen_media_kwargs=None, **kwargs):
+        if qwen_execution not in {"question", "shared_state"}:
+            raise ValueError("qwen_execution must be question or shared_state")
+        self.valen_head = valen_head or {"head_type": "mixer"}
+        self.qwen_execution = qwen_execution
+        self.valen_max_length = valen_max_length
+        self.valen_media_kwargs = valen_media_kwargs or {}
+        super().__init__(**kwargs)

unmerged/data_types.py

@@ -0,0 +1,12 @@
+"""Model-independent decision metadata; encoded tensors stay with each compiler."""
+from dataclasses import dataclass
+
+
+@dataclass
+class QuestionSpec:
+    qid: str
+    kind: str
+    keys: list
+    descriptions: list
+    target: list | None
+    is_text: bool

unmerged/export_manifest.json

@@ -0,0 +1,68 @@
+{
+  "format": "valen_qwen_unmerged_v1",
+  "created_utc": "2026-10-06T06:53:45.813755+00:00",
+  "model": "Valen-4B",
+  "source_checkpoint_sha256": "b4193b4e907f44b913fd54a8d6b998e7050a1528306269a000f1e5e8bd8dffc7",
+  "training_step": 325,
+  "training_records": 1225000,
+  "training_data_sha256": "648cb127fe4830f9faaa6d4595445cc2573ef0c152f8f74c9c776dc8a9223d7f",
+  "base": {
+    "repo": "Qwen/Qwen3.5-4B",
+    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+    "weight_sha256": {
+      "model.safetensors-00001-of-00002.safetensors": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
+      "model.safetensors-00002-of-00002.safetensors": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188"
+    },
+    "files_sha256": {
+      ".gitattributes": "34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930",
+      "LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
+      "README.md": "1406be1b6b8fd8a6545870da516912804756593628a1d0fb0a7965211e82a7bb",
+      "chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+      "config.json": "ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670",
+      "merges.txt": "a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d",
+      "model.safetensors-00001-of-00002.safetensors": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
+      "model.safetensors-00002-of-00002.safetensors": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188",
+      "model.safetensors.index.json": "cf3f798ee02ba45f9622aa8892a47369ab667d0afbf154ee7c2212de42e6302d",
+      "preprocessor_config.json": "27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516",
+      "tokenizer.json": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42",
+      "tokenizer_config.json": "316230d6a809701f4db5ea8f8fc862bc3a6f3229c937c174e674ff3ca0a64ac8",
+      "video_preprocessor_config.json": "7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13",
+      "vocab.json": "ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003"
+    },
+    "verified_from": "download_manifest.json; every LFS SHA256 matched"
+  },
+  "execution": "shared_state",
+  "trainable_tensor_count": 535,
+  "versions": {
+    "torch": "2.6.0",
+    "transformers": "5.4.0",
+    "peft": "0.18.1",
+    "safetensors": "0.6.2"
+  },
+  "files_sha256": {
+    "LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
+    "README.md": "19dade0ed4951da63d3f556939f52d16e81d6a77fea4372816754824c52a3983",
+    "batching.py": "ea45d4724bcafca559fc5a2a18092918b633746771c19da6554c29bc5f2cef64",
+    "chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
+    "compiler.py": "f35bab2e71b6beb1228daaa5b7baecdacf04eaf713aee14e8082b97090f910b4",
+    "config.json": "63d39d4399bf06747f5f06d6b68db7b17fa3e96e8bc219fb9823742a700a06c1",
+    "configuration_valen.py": "735b02402fb5b6c7a1652f6691ba66d066ef99e84500042faec396ff7a668fe5",
+    "data_types.py": "b06ce08081d8f7b2bf36abf302410a96086395e75a04f9e9b1c591db519b1641",
+    "heads.py": "63b5970a8bfa121eff94d50ffd35e1fa8387f3af5abccc5f5b75381fb084fc34",
+    "lora/README.md": "761f9a02d12a6434a0444d774b01bd6e13cd98c392c154f2442665b8e632819b",
+    "lora/adapter_config.json": "83e2bc08b862c2cec7326c9f1ef2f44580e422c5f4ba2eea1148c8cd12bc979a",
+    "lora/adapter_model.safetensors": "b162b622b6d92720a87c678931a88ff05fd5e4b93b356997153c124bfac87708",
+    "mixer.safetensors": "64d368877bac6f4d9e8662d07ed6f21d3e4e58bec4e7a5dc6b4ac555fb4028d9",
+    "modeling_valen.py": "07f87481334cb1e28e9daee036c27bab3654645abf3df7983346b316e7788848",
+    "modeling_valen_lora.py": "0102d4e71596dc7e1f6f3c95e6cf5fb514c50bff9f27c4870f16ad618874d4be",
+    "processor_config.json": "d89ef49ce9cd37fbf510158e13c1ef063d9286411c1ec9049932dbe0487143b1",
+    "responses.py": "298016b0070f1000a38d033e1048e038ff5df6fc1dcbe830601105b181f1d228",
+    "runtime_model.py": "98bf0443ca196900b9f2b056129c84932041b8b53395d111f2fe6d3f2bc1e0a6",
+    "schema.py": "253825862b60016e37fd1040492008ad7f303679c092313d8a6527cf509f3b6e",
+    "tokenizer.json": "87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4",
+    "tokenizer_config.json": "e98f1901ac6f0adff67b1d540bfa0c36ac1a0cf59eb72ed78146ef89aafa1182",
+    "validation.json": "6e66db946145cba9c89033b82036e048845ac0f4e81920dcc9a8992a7f88edd2",
+    "visual_merger.safetensors": "8379c8434e846ff5007de471cd6174092ac0d30699007fbea7780882106774c1",
+    "fresh_process_validation.json": "d5f7b9a17d0f1011d02177a0dc0af7530912c2a5f191f06ee98a3a94cbd39b0c"
+  }
+}

unmerged/fresh_process_validation.json

@@ -0,0 +1,113 @@
+{
+  "passed": true,
+  "isolated_python": true,
+  "valen_imports_blocked": true,
+  "lora_remains_unmerged": true,
+  "all_real_modalities_and_question_types_passed": true,
+  "sdpa_example_passed": true,
+  "cases": 20,
+  "max_probability_abs_error": 0.0,
+  "max_logit_abs_error": 0.0,
+  "details": [
+    {
+      "case": "shared_state/text_multi/choice",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/text_multi/noul",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/text_multi/score",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/text_multi/score_many",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/image_score/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/image_noul/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/image_choice/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/video_multi/q000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/video_multi/score_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "shared_state/video_multi/noul_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/choice",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/noul",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/score",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/text_multi/score_many",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/image_score/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/image_noul/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/image_choice/decision",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/video_multi/q000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/video_multi/score_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    },
+    {
+      "case": "question/video_multi/noul_000",
+      "probability_max_abs": 0.0,
+      "logit_max_abs": 0.0
+    }
+  ]
+}

unmerged/heads.py

@@ -0,0 +1,143 @@
+"""可替换的候选评分头。 / Interchangeable, candidate-wise decision heads."""
+import math
+
+import torch
+from torch import nn
+
+
+def head_signature(config):
+    """用于构建与恢复校验。 / Canonical architecture settings for loading."""
+    kind = config.get("head_type", "bilinear")
+    defaults = {
+        "bilinear": {"projection_dim": 256},
+        "mlp": {"head_hidden_dim": 512, "head_bottleneck_dim": 128},
+        "role_mlp": {"head_width": 256, "head_hidden_dim": 1024},
+        "mixer": {"head_width": 256, "head_layers": 2,
+                  "head_token_hidden_dim": 16, "head_channel_hidden_dim": 1024},
+    }
+    if kind not in defaults:
+        raise ValueError(f"Unknown Qwen head_type: {kind!r}")
+    values = {key: config.get(key, value) for key, value in defaults[kind].items()}
+    if any(type(value) is not int or value <= 0 for value in values.values()):
+        raise ValueError("Decision head dimensions must be positive integers")
+    return {"head_type": kind, **values}
+
+
+class DecisionHead(nn.Module):
+    """原有参数名保持兼容。 / Preserve legacy bilinear parameter names."""
+    def __init__(self, hidden_size, projection_dim=256):
+        super().__init__()
+        self.decision = nn.Linear(hidden_size, projection_dim, bias=False)
+        self.candidate = nn.Linear(hidden_size, projection_dim, bias=False)
+        self.scale = math.sqrt(projection_dim)
+
+    def forward(self, hidden, positions, decision_position):
+        d = self.decision(hidden[0, decision_position].float())
+        c = self.candidate(hidden[0, positions].float())
+        return (c * d).sum(-1) / self.scale
+
+    def extract_features(self, hidden, branch):
+        return hidden[:, [*branch.candidate_positions, branch.decision_position], :]
+
+    def score_features(self, features):
+        return self(features, list(range(features.shape[1] - 1)), features.shape[1] - 1)
+
+
+class MLPHead(nn.Module):
+    """每个候选共享一个 MLP。 / One shared MLP for arbitrary candidate counts."""
+    def __init__(self, hidden_size, hidden_dim=512, bottleneck_dim=128):
+        super().__init__()
+        self.decision_norm = nn.LayerNorm(hidden_size)
+        self.candidate_norm = nn.LayerNorm(hidden_size)
+        self.network = nn.Sequential(nn.Linear(2 * hidden_size, hidden_dim), nn.GELU(),
+                                     nn.Linear(hidden_dim, bottleneck_dim), nn.GELU(),
+                                     nn.Linear(bottleneck_dim, 1))
+
+    def extract_features(self, hidden, branch):
+        c = hidden[0, branch.candidate_positions]
+        d = hidden[0, branch.decision_position].expand_as(c)
+        return torch.stack((d, c), dim=1)
+
+    def score_features(self, features):
+        d, c = features.float().unbind(1)
+        return self.network(torch.cat((self.decision_norm(d), self.candidate_norm(c)), -1)).squeeze(-1)
+
+
+def role_features(hidden, branch):
+    """先池化再投影，可缓存冻结特征。 / Pool before trainable projections."""
+    spans = (getattr(branch, "context_span", None), getattr(branch, "instruction_span", None))
+    candidates = getattr(branch, "candidate_spans", None)
+    if any(span is None for span in spans) or not candidates or len(candidates) != len(branch.candidate_positions):
+        raise ValueError("Role heads require compiler context, instruction and candidate spans")
+
+    def pool(span):
+        start, end = span
+        if not 0 <= start < end <= hidden.shape[1]:
+            raise ValueError(f"Invalid role span: {span}")
+        return hidden[0, start:end].float().mean(0)
+
+    context, instruction = (pool(span) for span in spans)
+    decision = hidden[0, branch.decision_position].float()
+    return torch.stack([torch.stack((context, instruction, pool(span), decision)) for span in candidates])
+
+
+class RoleHead(nn.Module):
+    def __init__(self, hidden_size, width):
+        super().__init__()
+        self.input_norm = nn.LayerNorm(hidden_size)
+        self.projection = nn.Linear(hidden_size, width)
+        self.roles = nn.Parameter(torch.zeros(1, 4, width))
+
+    def extract_features(self, hidden, branch):
+        return role_features(hidden, branch)
+
+    def project(self, features):
+        return self.projection(self.input_norm(features.float())) + self.roles
+
+
+class RoleMLPHead(RoleHead):
+    """与 Mixer 使用相同输入。 / Control for the Mixer's richer input features."""
+    def __init__(self, hidden_size, width=256, hidden_dim=1024):
+        super().__init__(hidden_size, width)
+        self.network = nn.Sequential(nn.Linear(4 * width, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, 1))
+
+    def score_features(self, features):
+        return self.network(self.project(features).flatten(1)).squeeze(-1)
+
+
+class MixerBlock(nn.Module):
+    def __init__(self, width, token_hidden, channel_hidden):
+        super().__init__()
+        self.token_norm = nn.LayerNorm(width)
+        self.token_mlp = nn.Sequential(nn.Linear(4, token_hidden), nn.GELU(), nn.Linear(token_hidden, 4))
+        self.channel_norm = nn.LayerNorm(width)
+        self.channel_mlp = nn.Sequential(nn.Linear(width, channel_hidden), nn.GELU(), nn.Linear(channel_hidden, width))
+
+    def forward(self, x):
+        # 混合角色，不混合不同候选。 / Mix fixed roles, never candidate rows.
+        x = x + self.token_mlp(self.token_norm(x).transpose(1, 2)).transpose(1, 2)
+        return x + self.channel_mlp(self.channel_norm(x))
+
+
+class MixerHead(RoleHead):
+    def __init__(self, hidden_size, width=256, layers=2, token_hidden=16, channel_hidden=1024):
+        super().__init__(hidden_size, width)
+        self.blocks = nn.Sequential(*(MixerBlock(width, token_hidden, channel_hidden) for _ in range(layers)))
+        self.output = nn.Sequential(nn.LayerNorm(width), nn.Linear(width, 1))
+
+    def score_features(self, features):
+        # 第三行是候选角色。 / Read the updated candidate role.
+        return self.output(self.blocks(self.project(features))[:, 2]).squeeze(-1)
+
+
+def build_head(hidden_size, config):
+    settings = head_signature(config)
+    kind = settings.pop("head_type")
+    if kind == "bilinear":
+        return DecisionHead(hidden_size, settings["projection_dim"])
+    if kind == "mlp":
+        return MLPHead(hidden_size, settings["head_hidden_dim"], settings["head_bottleneck_dim"])
+    if kind == "role_mlp":
+        return RoleMLPHead(hidden_size, settings["head_width"], settings["head_hidden_dim"])
+    return MixerHead(hidden_size, settings["head_width"], settings["head_layers"],
+                     settings["head_token_hidden_dim"], settings["head_channel_hidden_dim"])

unmerged/lora/README.md

@@ -0,0 +1,206 @@
+---
+base_model: ''
+library_name: peft
+tags:
+- 'base_model:adapter:'
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.18.1

unmerged/lora/adapter_config.json

@@ -0,0 +1,54 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "Qwen3_5TextModel",
+    "parent_library": "transformers.models.qwen3_5.modeling_qwen3_5"
+  },
+  "base_model_name_or_path": "Qwen/Qwen3.5-4B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+  "target_modules": [
+    "in_proj_qkv",
+    "out_proj",
+    "in_proj_z",
+    "gate_proj",
+    "o_proj",
+    "up_proj",
+    "k_proj",
+    "in_proj_a",
+    "in_proj_b",
+    "v_proj",
+    "q_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

unmerged/lora/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b162b622b6d92720a87c678931a88ff05fd5e4b93b356997153c124bfac87708
+size 259784032

unmerged/mixer.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64d368877bac6f4d9e8662d07ed6f21d3e4e58bec4e7a5dc6b4ac555fb4028d9
+size 6867012

unmerged/modeling_valen.py

@@ -0,0 +1,84 @@
+"""Complete merged Qwen + decision head. / 已合并的完整 Qwen 和决策头。"""
+import torch
+from transformers import AutoProcessor
+from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5Model, Qwen3_5PreTrainedModel
+
+from .configuration_valen import ValenQwenConfig
+from .heads import build_head
+from .runtime_model import ValenQwen
+from .compiler import Compiler
+from .responses import answer
+# List transitive files explicitly for Transformers' local dynamic loader.
+# 本地动态加载需要显式列出间接依赖，否则搬移后可能缺少这些模块。
+from .batching import collate_branches
+from .schema import candidates
+from .data_types import QuestionSpec
+
+
+class ValenQwenForDecisionMaking(Qwen3_5PreTrainedModel):
+    config_class = ValenQwenConfig
+    base_model_prefix = "backbone"
+    model_name = "Valen"
+    architecture = "qwen"
+    # The trained Mixer stays FP32 with a BF16 backbone. / BF16 基座、FP32 决策头。
+    _keep_in_fp32_modules_strict = ["head"]
+
+    def __init__(self, config, backbone=None, head=None):
+        super().__init__(config)
+        if (backbone is None) != (head is None):
+            raise ValueError("Supply both a merged backbone and a trained head")
+        self.backbone = backbone if backbone is not None else Qwen3_5Model(config)
+        self.head = head if head is not None else build_head(config.text_config.hidden_size, config.valen_head)
+        self.head.float()
+        self._processor = None
+        # Supplied weights are already trained: do not reinitialize them.
+        # 传入的参数已训练完成，不能再次调用初始化覆盖它们。
+        if backbone is not None:
+            for module in self.modules():
+                module._is_hf_initialized = True
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.backbone.set_input_embeddings(value)
+
+    def forward(self, question=None, **inputs):
+        if question is None:
+            return self.backbone(**inputs)
+        if hasattr(question, "questions"):
+            return self.forward_state(question)
+        return ValenQwen.forward(self, question)
+
+    # Use the exact trained readouts and batching implementation.
+    # 保留训练时的特征读取和 batch 逻辑。
+    extract_features = ValenQwen.extract_features
+    extract_state_features = ValenQwen.extract_state_features
+    _state_features = ValenQwen._state_features
+    forward_state = ValenQwen.forward_state
+    forward_state_batch = ValenQwen.forward_state_batch
+    forward_batch = ValenQwen.forward_batch
+    score_features = ValenQwen.score_features
+
+    def make_compiler(self, media_root=".", execution=None, processor=None):
+        if processor is None:
+            if self._processor is None:
+                self._processor = AutoProcessor.from_pretrained(self.config._name_or_path, trust_remote_code=True)
+            processor = self._processor
+        return Compiler(processor, media_root, self.config.valen_max_length, self.config.valen_media_kwargs,
+                        execution=execution or self.config.qwen_execution)
+
+    @torch.inference_mode()
+    def predict(self, request, *, media_root=".", execution=None, temperature=1.0, processor=None):
+        self.eval()
+        # Only request content reaches the compiler. / 只编译输入，不读取标签。
+        request = request["request"] if "request" in request else request
+        compiler = self.make_compiler(media_root, execution, processor)
+        state = compiler.compile({"request": request})
+        logits = self.forward_state(state) if state.inputs is not None else [self(q) for q in state.questions]
+        answers = {q.qid: answer(q, z, temperature[q.kind] if isinstance(temperature, dict) else temperature)
+                   for q, z in zip(state.questions, logits)}
+        return {"model": self.model_name, "answers": answers,
+                "usage": {"input_tokens": state.logical_tokens, "output_tokens": 0},
+                "internal_usage": {"compute_tokens": state.compute_tokens}}

unmerged/modeling_valen_lora.py

@@ -0,0 +1,92 @@
+"""Original base + separate trained deltas. / 原始基座与未融合的训练权重。"""
+import hashlib
+from pathlib import Path
+
+import torch
+from peft import PeftModel
+from safetensors.torch import load_file
+from transformers import Qwen3_5ForConditionalGeneration
+
+from .configuration_valen import ValenQwenConfig
+from .modeling_valen import ValenQwenForDecisionMaking
+# Explicit dependencies keep local HF dynamic loading portable.
+# 显式列出依赖，保证 HF 本地动态加载能复制完整的运行代码。
+from .heads import build_head
+from .compiler import Compiler
+from .runtime_model import ValenQwen
+from .responses import answer
+from .batching import collate_branches
+from .schema import candidates
+from .data_types import QuestionSpec
+
+
+def _verify_local_base(base_path, expected):
+    if not Path(base_path).is_dir():
+        return
+    for filename, digest in expected.items():
+        path = Path(base_path) / filename
+        if not path.is_file():
+            raise FileNotFoundError(f"Original base shard is required: {path}")
+        actual = hashlib.sha256()
+        with path.open("rb") as stream:
+            for chunk in iter(lambda: stream.read(8 << 20), b""):
+                actual.update(chunk)
+        if actual.hexdigest() != digest:
+            raise ValueError(f"Base weights differ from the trained revision: {filename}")
+
+
+class ValenQwenWithLoRA(ValenQwenForDecisionMaking):
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, config=None,
+                        base_model_path=None, device=None, dtype=None, attn_implementation=None,
+                        is_trainable=False, local_files_only=True, output_loading_info=False,
+                        verify_base=True, **kwargs):
+        if model_args:
+            raise TypeError("Pass loader options as named arguments")
+        folder = Path(pretrained_model_name_or_path).resolve(strict=True)
+        if config is None:
+            config = ValenQwenConfig.from_pretrained(folder, local_files_only=True)
+        base = str(base_model_path or config.valen_base["repo"])
+        dtype = dtype or kwargs.pop("torch_dtype", None) or config.dtype or torch.bfloat16
+        if isinstance(dtype, str):
+            dtype = getattr(torch, dtype)
+        if dtype != torch.bfloat16:
+            raise ValueError("Use dtype=torch.bfloat16 to reproduce this adapter bundle")
+        attention = attn_implementation or config._attn_implementation or "sdpa"
+        if kwargs.pop("device_map", None) is not None:
+            raise ValueError("Use device=... or model.to(...) for this adapter bundle")
+        # AutoModel also supplies flags for loading the bundle's custom code.
+        # AutoModel 会传入自定义代码标记，它们不应改变原始基座的加载配置。
+        for name in ("trust_remote_code", "_from_auto", "_commit_hash", "adapter_kwargs", "revision", "subfolder"):
+            kwargs.pop(name, None)
+        base_options = {name:kwargs.pop(name) for name in
+                        ("cache_dir", "token", "force_download", "proxies", "low_cpu_mem_usage") if name in kwargs}
+        if kwargs:
+            raise TypeError(f"Unsupported loader options: {sorted(kwargs)}")
+        if verify_base:
+            _verify_local_base(base, config.valen_base["weight_sha256"])
+        # Load the published conditional model first to preserve language prefixes.
+        # 先完整加载官方模型，再取出主干，避免语言参数前缀被错误映射。
+        container, loading = Qwen3_5ForConditionalGeneration.from_pretrained(
+            base, revision=config.valen_base["revision"], dtype=dtype,
+            attn_implementation=attention, local_files_only=local_files_only,
+            output_loading_info=True, **base_options)
+        if any(loading.get(key) for key in ("missing_keys", "unexpected_keys", "mismatched_keys", "error_msgs")):
+            raise ValueError(f"Incomplete original base load: {loading}")
+        backbone = container.model
+        del container
+        backbone.requires_grad_(False)
+        backbone.language_model = PeftModel.from_pretrained(
+            backbone.language_model, folder / "lora", is_trainable=is_trainable,
+            autocast_adapter_dtype=True, torch_device="cpu", local_files_only=True)
+        backbone.visual.merger.load_state_dict(load_file(folder / "visual_merger.safetensors"), strict=True)
+        head = build_head(config.text_config.hidden_size, config.valen_head).float()
+        head.load_state_dict(load_file(folder / "mixer.safetensors"), strict=True)
+        config._name_or_path = str(folder)
+        model = cls(config, backbone=backbone, head=head)
+        model.head.requires_grad_(is_trainable)
+        model.backbone.visual.merger.requires_grad_(is_trainable)
+        if device is not None:
+            model.to(device)
+        model.eval()
+        return (model, loading) if output_loading_info else model

unmerged/processor_config.json

@@ -0,0 +1,60 @@
+{
+  "image_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "merge_size": 2,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "size": {
+      "longest_edge": 16777216,
+      "shortest_edge": 65536
+    },
+    "temporal_patch_size": 2
+  },
+  "processor_class": "Qwen3VLProcessor",
+  "video_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "do_sample_frames": true,
+    "fps": 2,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_frames": 768,
+    "merge_size": 2,
+    "min_frames": 4,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "return_metadata": false,
+    "size": {
+      "longest_edge": 25165824,
+      "shortest_edge": 4096
+    },
+    "temporal_patch_size": 2,
+    "video_processor_type": "Qwen3VLVideoProcessor"
+  }
+}

unmerged/responses.py

@@ -0,0 +1,29 @@
+import math
+import torch
+
+def answer(question, logits, temperature=1.0):
+    if not math.isfinite(temperature) or temperature <= 0:
+        raise ValueError("Calibration temperature must be finite and positive")
+    if logits.ndim != 1 or len(logits) != len(question.keys) or not torch.isfinite(logits).all():
+        raise ValueError("Expected one finite logit per candidate")
+    probs = (logits.float() / temperature).softmax(-1).detach().cpu().tolist()
+    total = sum(probs)
+    probs = [p / total for p in probs]
+    result = {"type": question.kind}
+    if question.kind == "noul":
+        result["noul"] = probs[question.keys.index("true")]
+        return result
+    result["probabilities"] = dict(zip(question.keys, probs))
+    mode = max(range(len(probs)), key=probs.__getitem__)
+    if question.kind == "choice":
+        result["choice"] = question.keys[mode]
+        result["confidence"] = 1.0 if len(probs) == 1 else max(0.0, (max(probs) - 1/len(probs)) / (1 - 1/len(probs)))
+    else:
+        result["score"] = sum(i * p for i, p in enumerate(probs))
+        result["legend"] = dict(zip(question.keys, question.descriptions))
+        # Independently implemented formula from pinned system-one adapter:
+        # fb52b1030b7fc1f4f1cf39910afa5da54f9835e3, confidence_metrics.py.
+        distance = sum(p * abs(i - mode) for i, p in enumerate(probs))
+        uniform_mad = sum(abs(i - (len(probs)-1)/2) for i in range(len(probs))) / len(probs)
+        result["confidence"] = max(0.0, 1.0 - distance / uniform_mad)
+    return result

unmerged/runtime_model.py

@@ -0,0 +1,101 @@
+import torch
+from torch import nn
+MODEL_NAME = "Valen"
+from .heads import DecisionHead, build_head
+from .batching import branch_batches, collate_branches
+
+
+class ValenQwen(nn.Module):
+    """Qwen3.5 backbone with a shared candidate decision head."""
+
+    model_name = MODEL_NAME
+    architecture = "qwen"
+
+    def __init__(self, backbone, projection_dim=256, head_config=None):
+        super().__init__()
+        self.backbone = backbone
+        self.head = build_head(backbone.config.text_config.hidden_size,
+                               dict({"projection_dim": projection_dim}, **(head_config or {})))
+
+    def forward(self, question):
+        device = next(self.backbone.parameters()).device
+        outputs = []
+        for branch in question.branches:
+            inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in branch.inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            outputs.append(self.head.score_features(self.head.extract_features(hidden, branch)))
+        return torch.cat(outputs)
+
+    def extract_features(self, question):
+        """缓存投影前的特征。 / Cache raw readouts before trainable head layers."""
+        device = next(self.backbone.parameters()).device
+        features = []
+        for branch in question.branches:
+            inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in branch.inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            features.append(self.head.extract_features(hidden, branch))
+        return features
+
+    def extract_state_features(self, state):
+        """一次编码 state，读取各题特征。 / One backbone graph for every question."""
+        if not state.questions:
+            return []
+        if state.inputs is None:
+            raise ValueError("Shared-state features require compiled shared inputs")
+        device = next(self.backbone.parameters()).device
+        inputs = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in state.inputs.items()}
+        hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+        return self._state_features(hidden, state)
+
+    def _state_features(self, hidden, state):
+        return [[self.head.extract_features(hidden, branch) for branch in q.branches]
+                for q in state.questions]
+
+    def forward_state(self, state):
+        return [self.score_features(features) for features in self.extract_state_features(state)]
+
+    def forward_state_batch(self, states, max_tokens=32768):
+        """不同 state 组成 batch，同一 state 的题目共用一行。 / One row per shared state."""
+        device = next(self.backbone.parameters()).device
+        pad_token_id = getattr(self.backbone.config.text_config, "pad_token_id", None) or 0
+        outputs = [[] for _ in states]
+        representatives = []
+        indices = []
+        for index, state in enumerate(states):
+            if not state.questions:
+                continue
+            if state.inputs is None or any(branch.inputs is not state.inputs
+                    for q in state.questions for branch in q.branches):
+                raise ValueError("Shared-state batch requires one shared input per state")
+            representatives.append(state.questions[0])
+            indices.append(index)
+        for batch in branch_batches(representatives, max_tokens):
+            inputs = collate_branches([branch for _, branch in batch], pad_token_id)
+            inputs = {key: value.to(device) for key, value in inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            for row, (index, _) in enumerate(batch):
+                original = indices[index]
+                features = self._state_features(hidden[row:row + 1], states[original])
+                outputs[original] = [self.score_features(f) for f in features]
+        return outputs
+
+    def forward_batch(self, questions, max_tokens=32768):
+        """Forward several QAs together; each keeps its own decision head inputs.
+
+        多条 QA 并行；每条保留独立的候选、角色区间和 Score 分支。
+        """
+        device = next(self.backbone.parameters()).device
+        pad_token_id = getattr(self.backbone.config.text_config, "pad_token_id", None) or 0
+        outputs = [[] for _ in questions]
+        for batch in branch_batches(questions, max_tokens):
+            inputs = collate_branches([branch for _, branch in batch], pad_token_id)
+            inputs = {key: value.to(device) for key, value in inputs.items()}
+            hidden = self.backbone(**inputs, use_cache=False, return_dict=True).last_hidden_state
+            for row, (question_index, branch) in enumerate(batch):
+                features = self.head.extract_features(hidden[row:row + 1], branch)
+                outputs[question_index].append(self.head.score_features(features))
+        return [torch.cat(branches) for branches in outputs]
+
+    def score_features(self, features, head=None):
+        head = self.head if head is None else head
+        return torch.cat([head.score_features(hidden) for hidden in features])

unmerged/schema.py

@@ -0,0 +1,71 @@
+"""Strict JSONL contract. IDs and labels never enter the model prompt."""
+import json
+import math
+from pathlib import Path
+
+
+def candidates(question):
+    kind = question.get("type")
+    if not isinstance(question.get("instructions"), str) or not question["instructions"].strip():
+        raise ValueError("instructions must be nonempty text")
+    criteria = question.get("criteria")
+    if kind == "noul":
+        return [("true", "True / 是：满足问题中的条件。"), ("false", "False / 否：不满足问题中的条件。")]
+    if kind == "choice":
+        if not isinstance(criteria, dict) or not 1 <= len(criteria) <= 255:
+            raise ValueError("Choice needs 1–255 named criteria")
+        result = list(criteria.items())
+    elif kind == "score":
+        if not isinstance(criteria, list) or not 2 <= len(criteria) <= 255:
+            raise ValueError("Score criteria must be an ordered list of 2–255 descriptions")
+        result = [(str(i), text) for i, text in enumerate(criteria)]
+    else:
+        raise ValueError(f"Unsupported question type: {kind}")
+    if any(not isinstance(k, str) or not k or not isinstance(v, str) or not v.strip() for k, v in result):
+        raise ValueError("Candidate names and descriptions must be nonempty strings")
+    return result
+
+
+def target_distribution(target, keys):
+    if target is None or target.get("probabilities") is None:
+        return None
+    probs = target["probabilities"]
+    if not isinstance(probs, dict) or set(probs) != set(keys):
+        raise ValueError(f"Probability keys must exactly match {keys}")
+    values = [probs[k] for k in keys]
+    if any(isinstance(v, bool) or not isinstance(v, (int, float)) or not math.isfinite(v) or v < 0 for v in values):
+        raise ValueError("Probabilities must be finite nonnegative numbers")
+    if not math.isclose(sum(values), 1.0, abs_tol=1e-6):
+        raise ValueError("Target probabilities must sum to 1")
+    return values
+
+
+def validate_record(record, candidate_fn=candidates):
+    request = record["request"]
+    state = request["state"]
+    if not isinstance(state, (str, dict)):
+        raise ValueError("state must be text or a messages object")
+    if not isinstance(request["questions"], dict) or not request["questions"]:
+        raise ValueError("questions must be a nonempty mapping")
+    targets = record.get("targets", {})
+    if set(targets) - set(request["questions"]):
+        raise ValueError("Unknown target question IDs")
+    for qid, question in request["questions"].items():
+        target_distribution(targets.get(qid), [k for k, _ in candidate_fn(question)])
+    if not record.get("group_id"):
+        raise ValueError("group_id is required for leakage-free splits")
+    return record
+
+
+def read_jsonl(path, candidate_fn=candidates):
+    records = []
+    with Path(path).open(encoding="utf-8") as stream:
+        for line_number, line in enumerate(stream, 1):
+            if line.strip():
+                try:
+                    records.append(validate_record(json.loads(line), candidate_fn))
+                except (ValueError, KeyError, TypeError, AttributeError) as exc:
+                    raise ValueError(f"{path}:{line_number}: {exc}") from exc
+    if not records:
+        raise ValueError("Empty dataset")
+    return records

unmerged/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4
+size 19989343

unmerged/tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "processor_class": "Qwen3VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

unmerged/validation.json

@@ -0,0 +1,136 @@
+{
+  "passed": true,
+  "all_training_deltas_bitwise_reloaded": true,
+  "all_backbone_buffers_bitwise_preserved": true,
+  "trainable_tensor_count": 535,
+  "reload_vs_checkpoint": {
+    "cases": 20,
+    "changed_labels": [],
+    "max_probability_abs_error": 0.0,
+    "max_logit_abs_error": 0.0,
+    "details": [
+      {
+        "case": "shared_state/text_multi/choice",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/text_multi/noul",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/text_multi/score",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/text_multi/score_many",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/image_score/decision",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/image_noul/decision",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/image_choice/decision",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/video_multi/q000",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/video_multi/score_000",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "shared_state/video_multi/noul_000",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/text_multi/choice",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/text_multi/noul",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/text_multi/score",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/text_multi/score_many",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/image_score/decision",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/image_noul/decision",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/image_choice/decision",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/video_multi/q000",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/video_multi/score_000",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      },
+      {
+        "case": "question/video_multi/noul_000",
+        "probability_max_abs": 0.0,
+        "logit_max_abs": 0.0,
+        "label_changed": false
+      }
+    ]
+  },
+  "public_predict_passed": true,
+  "scope": "Text/image/video; question/shared_state; Choice/Noul/Score, including 12-level Score."
+}