--- library_name: transformers license: apache-2.0 base_model: Qwen/Qwen3-8B-Base tags: - roleplay - conversational - axolotl - qwen --- # 残响 Qwen3 8B(第一系列) [English](./README.md) | [简体中文](./README-cn.md) *空气中飘浮着一缕尘埃。它仿佛来自某个逝去的时代,但你无从追溯。它落在你的舌尖,滋味奇妙。* ![image/png](https://cdn-uploads.huggingface.co/production/uploads/634262af8d8089ebaefd410e/_ovgodU331FO4YAqFGCnk.png) 「残响」是一系列专注于SFW与NSFW角色扮演及对话的微调大语言模型。 ## 量化版本 GGUF: - 待补充! EXL3: - 待补充! EXL2: - 待补充! 其他格式: - 待补充! ## 推荐参数 对话模板: ChatML 采样器设置: - 温度值 `0.8` - 最小概率阈值 `0.1` - 存在惩罚 `0.5` ## 致谢 特别感谢Allura和ilya <3 衷心感谢以下项目的开发者: - Axolotl(训练框架) - 通义千问/Qwen/阿里巴巴(基础模型) - Prime Intellect(算力支持) - 以及我的银行(资金支持) ## 其他信息 [Built with Axolotl](https://github.com/axolotl-ai-cloud/axolotl)
查看Axolotl配置 axolotl版本: `0.10.0.dev0` ```yaml # === 模型配置 === base_model: Qwen/Qwen3-8B-Base load_in_8bit: false load_in_4bit: false # === 训练设置 === num_epochs: 2 micro_batch_size: 32 gradient_accumulation_steps: 1 sequence_len: 8192 sample_packing: true pad_to_sequence_len: true # === 超参数配置 === optimizer: apollo_adamw_layerwise # Apollo-mini配置: optim_args: "proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200" # 标准Apollo配置: # optim_args: optim_target_modules: all_linear learning_rate: 2e-5 lr_scheduler: rex weight_decay: 0.01 warmup_ratio: 0 # === 数据配置 === datasets: - path: allura-org/inkmix-v3.0 type: chat_template split: train field_messages: conversations message_field_role: from message_field_content: value dataset_prepared_path: last_run_prepared chat_template: chatml # === 插件 === plugins: - axolotl.integrations.liger.LigerPlugin - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin # === 硬件优化 === gradient_checkpointing: unsloth gradient_checkpointing_kwargs: use_reentrant: false liger_rope: true liger_rms_norm: true liger_glu_activation: true cut_cross_entropy: true # === Wandb追踪 === wandb_project: qwen3-8b-inkmix-v3 # === 检查点 === saves_per_epoch: 2 save_total_limit: 3 # === 高级设置 === output_dir: /ephemeral/ckpts bf16: auto flash_attention: true train_on_inputs: false group_by_length: false logging_steps: 1 trust_remote_code: true ```