lordx64/reasoning-distill-opus-4-7-max-sft

--- license: apache-2.0 language: - en tags: - reasoning - chain-of-thought - distillation - claude - opus-4-7 - sft - qwen-chat-template task_categories: - text-generation size_categories: - 1K<n<10K dataset_info: features: - name: text dtype: string splits: - name: train num_bytes: 29328233 num_examples: 7823 download_size: 15809651 dataset_size: 29328233 configs: - config_name: default data_files: - split: train path: data/train-* --- # Reasoning traces from Claude Opus 4.7 — SFT-ready 7,823 single-turn reasoning conversations from **Claude Opus 4.7** reformatted for supervised fine-tuning with `trl.SFTTrainer` + `train_on_responses_only`. Each row is a single `text` field containing a full Qwen-style chat-template conversation. ## Provenance Every conversation's assistant response (including the `<think>...</think>` block) is output from **`claude-opus-4-7`** with Anthropic's `extended-thinking` enabled. This is the SFT-reformatted version of the raw dataset: - **Raw upstream**: [`lordx64/reasoning-distill-claude-opus-4-7-max`](https://huggingface.co/datasets/lordx64/reasoning-distill-claude-opus-4-7-max) — has `model`, `thinking`, `response`, and `source_dataset` columns. Check there for full attribution. ### Why this dataset has `4-7` in the name but sources mention 4.6 The *prompts* were reused from earlier distillation corpora (some of which have "4.6" in their names because they originally targeted Opus 4.6). The *responses* in this dataset are all regenerated from scratch against Opus 4.7 — which is what determines the dataset's name. See the [raw dataset card](https://huggingface.co/datasets/lordx64/reasoning-distill-claude-opus-4-7-max) for the full prompt→response pipeline. ## Format Each `text` value is a complete chat conversation in Qwen chat template with thinking: ``` <|im_start|>system {system_prompt}<|im_end|> <|im_start|>user {user_prompt}<|im_end|> <|im_start|>assistant <think> {opus_4_7_extended_thinking} </think> {opus_4_7_final_answer}<|im_end|> ``` Ready to feed to `SFTTrainer` with `dataset_text_field="text"`. The model we trained uses `train_on_responses_only` to mask loss on the user/system side — gradients only flow through the assistant turn, including its thinking tokens. ## Size - **Rows**: 7,823 (a few dropped from the raw 8,124 during formatting — rows where `stop_reason != end_turn` or where `thinking` / `response` was empty) - **Avg tokens per row**: ~4k (Qwen3 tokenizer), with long-tail reasoning chains going up to 32k tokens ## Model trained on this dataset [`lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled`](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled) — attention-only LoRA, r=16, 2 epochs, single H200. Preliminary evals: GSM8K 84.3%, MMLU-Pro 74.9%. ## Terms of use Generated using Anthropic's Claude Opus 4.7 via the official API. Downstream users should confirm compliance with [Anthropic's usage policies](https://www.anthropic.com/legal/usage-policy) for their specific use case. License: Apache 2.0 (for the dataset packaging; content itself is subject to the upstream terms above).许可证: Apache-2.0 语言: - 英语 标签: - 推理（reasoning） - 思维链（chain-of-thought） - 知识蒸馏（distillation） - Claude - Opus-4-7 - 监督微调（Supervised Fine-Tuning，SFT） - Qwen对话模板（qwen-chat-template） 任务类别: - 文本生成 样本规模: 1000 < 样本数 < 10000 数据集信息: 特征: - 字段名: text，数据类型: 字符串 数据划分: - 划分名: train，字节数: 29328233，样本数: 7823 下载大小: 15809651，数据集总大小: 29328233 配置项: - 配置名称: default，数据文件: - 训练划分: 路径为 data/train-* # Claude Opus 4.7 推理轨迹 — 适配监督微调（Supervised Fine-Tuning，SFT） 本数据集包含7823条来自Claude Opus 4.7的单轮推理对话，已针对使用`trl.SFTTrainer`与`train_on_responses_only`模式的监督微调进行格式重构。每条数据对应一个`text`字段，其中包含完整的Qwen风格对话模板会话。 ## 来源 所有对话的助手回复（包含`<think>...</think>`代码块）均由启用Anthropic官方`extended-thinking`（扩展思维）功能的`claude-opus-4-7`生成。本数据集为原始数据集的SFT格式重构版本： - **上游原始数据集**：[`lordx64/reasoning-distill-claude-opus-4-7-max`](https://huggingface.co/datasets/lordx64/reasoning-distill-claude-opus-4-7-max)，该数据集包含`model`、`thinking`、`response`与`source_dataset`字段，完整的署名信息可参考该数据集页面。 ### 为何本数据集名称中为`4-7`，但上游来源提及`4.6` 本数据集的**提示词（prompts）**复用自早期的知识蒸馏语料库（部分语料库名称中带有`4.6`，因其最初针对Opus 4.6构建），但本数据集内的**回复（responses）**均基于Opus 4.7从头生成，这也是本数据集命名的依据。完整的提示词→回复生成流程可参考[原始数据集卡片](https://huggingface.co/datasets/lordx64/reasoning-distill-claude-opus-4-7-max)。 ## 数据格式 每条`text`字段的值均为采用Qwen对话模板格式且包含思维过程的完整会话，格式如下： <|im_start|>system {system_prompt}<|im_end|> <|im_start|>user {user_prompt}<|im_end|> <|im_start|>assistant <think> {opus_4_7_extended_thinking} </think> {opus_4_7_final_answer}<|im_end|> 该格式可直接用于`SFTTrainer`，只需指定`dataset_text_field="text"`。我们训练的模型采用`train_on_responses_only`模式，对用户与系统侧的损失进行掩码，仅在助手回复轮次（包含思维标记）上计算梯度。 ## 数据集规模 - **数据条数**：7823条（原始数据集共8124条，在格式重构过程中移除了部分不符合要求的样本：即`stop_reason != end_turn`，或`thinking`/`response`字段为空的样本） - **单条样本平均令牌数**：约4000（基于Qwen3分词器），部分长尾推理链的令牌数可达32000。 ## 基于本数据集训练的模型 [`lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled`](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled)，采用仅注意力机制的LoRA（Low-Rank Adaptation，低秩适配）配置，秩大小r=16，训练2个epoch，使用单张H200显卡完成训练。初步评估结果为：GSM8K准确率84.3%，MMLU-Pro准确率74.9%。 ## 使用条款 本数据集通过Anthropic官方API使用Claude Opus 4.7生成。下游使用者需根据自身具体应用场景，确认符合[Anthropic使用政策](https://www.anthropic.com/legal/usage-policy)的相关要求。 许可证：Apache 2.0（仅针对数据集的封装格式；数据集内容本身需遵循上述上游条款）。