mayflowergmbh/boldt-dc-1b-orpo-onpolicy-de
收藏资源简介：
一个包含56,413行的德语偏好数据集，其中`rejected`字段是[`mayflowergmbh/boldt-dc-1b-german-it-16k`](https://huggingface.co/mayflowergmbh/boldt-dc-1b-german-it-16k) SFT模型在贪婪解码下的实际输出，而非现成的替代答案。其目的是用于On-policy ORPO/DPO训练：针对模型自身当前失败模式进行训练，通常比针对通用弱基线训练能提供更强烈的梯度信号，因为每个偏好对都记录了该特定模型当前行为(`rejected`)与我们期望行为(`chosen`)之间的具体差距。数据集由两个德语偏好数据集合并和统一而成：`johannhartmann/orpo-dpo-mix-40k-llama3-de`（保留43,947行）和`mayflowergmbh/intel_orca_dpo_pairs_de`（保留12,466行）。`prompt`字段已使用Boldt聊天标记（`<|system|>` / `<|user|>` / `<|assistant|>`）格式化，并以尾随的`<|assistant|> `结尾，因此无需应用聊天模板。数据集包含多个特征字段，如`row_id`、`source`、`prompt`、`chosen`、`rejected`、`original_rejected`、`prompt_tok_len`、`gen_model`、`gen_engine`、`gen_seed`、`gen_temperature`、`timestamp_utc`、`is_refusal`、`chosen_tok_len`、`rejected_tok_len`。数据集主要用于`mayflowergmbh/boldt-dc-1b-german-it-16k`及其相近衍生模型的ORPO/DPO/IPO/SimPO后训练，也可用于诊断该SFT模型的失败模式。数据集在生成后经过过滤，保留了具有真实梯度信号的行。长度分布显示`chosen`和`rejected`的单词数统计信息。`rejected`的拒绝率为1.2%（在`is_refusal`中标记）。数据集基于Apache 2.0许可证发布，但上游源数据集的许可证需在商业使用前审查。
This German preference dataset comprises 56,413 rows. The `rejected` field contains the actual outputs of the [`mayflowergmbh/boldt-dc-1b-german-it-16k`](https://huggingface.co/mayflowergmbh/boldt-dc-1b-german-it-16k) SFT model under greedy decoding, rather than pre-prepared alternative answers. Its primary purpose is for on-policy ORPO/DPO training: training against the model's own current failure modes typically yields stronger gradient signals than training against generic weak baselines, as each preference pair documents the specific gap between the model's current behavior (`rejected`) and the desired behavior (`chosen`). This dataset is consolidated and unified from two German preference datasets: `johannhartmann/orpo-dpo-mix-40k-llama3-de` (retaining 43,947 rows) and `mayflowergmbh/intel_orca_dpo_pairs_de` (retaining 12,466 rows). The `prompt` field has been formatted using Boldt chat tokens (`<|system|>` / `<|user|>` / `<|assistant|>`) and ends with a trailing `<|assistant|> `, so no additional chat template application is required. The dataset includes multiple feature fields, such as `row_id`, `source`, `prompt`, `chosen`, `rejected`, `original_rejected`, `prompt_tok_len`, `gen_model`, `gen_engine`, `gen_seed`, `gen_temperature`, `timestamp_utc`, `is_refusal`, `chosen_tok_len`, and `rejected_tok_len`. This dataset is primarily intended for post-training via ORPO/DPO/IPO/SimPO for `mayflowergmbh/boldt-dc-1b-german-it-16k` and its closely related derivative models, and can also be used to diagnose failure modes of the aforementioned SFT model. After generation, the dataset was filtered to retain only rows with valid gradient signals. The length distribution presents word count statistics for the `chosen` and `rejected` fields. The refusal rate for the `rejected` samples is 1.2%, as marked in the `is_refusal` field. This dataset is released under the Apache 2.0 license, but the licenses of the upstream source datasets must be reviewed prior to commercial usage.
