许可证:apache-2.0
语言:
标签:
流水线标签:文本生成
库名称:transformers
数据集:
- anthracite-org/c2_logs_32k_llama3_qwen2_v1.2_no_system
- anthracite-org/kalo-opus-instruct-22k-no-refusal-no-system
- anthracite-org/kalo-opus-instruct-3k-filtered-no-system
- anthracite-org/nopm_claude_writing_fixed
- anthracite-org/kalo_opus_misc_240827_no_system
- anthracite-org/kalo_misc_part2_no_system

本仓库包含模型的GGUF量化版本。如需原始权重,请访问此处。
该系列模型旨在复现Claude 3模型(特别是Sonnet和Opus)的文本生成质量。
本模型基于mistralai/Mistral-Nemo-Instruct-2407进行微调。
提示格式
典型输入格式如下:
<s>[INST] 系统消息
用户消息[/INST] 助手消息</s>[INST] 用户消息[/INST]
SillyTavern模板
以下是用于SillyTavern的指令模板和上下文模板。
上下文模板
默认SillyTavern模板即可适用
指令模板
默认SillyTavern模板即可适用
Axolotl配置
查看axolotl配置
基础模型: mistralai/Mistral-Nemo-Instruct-2407
模型类型: AutoModelForCausalLM
分词器类型: AutoTokenizer
中心模型ID: anthracite-org/magnum-v4-12b-r2
中心策略: "all_checkpoints"
推送数据集至中心:
使用认证令牌: true
插件:
- axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true
8位加载: false
4位加载: false
严格模式: false
数据集:
- 路径: anthracite-org/c2_logs_32k_llama3_qwen2_v1.2_no_system
类型: custommistralv3tekken
- 路径: anthracite-org/kalo-opus-instruct-22k-no-refusal-no-system
类型: custommistralv3tekken
- 路径: anthracite-org/kalo-opus-instruct-3k-filtered-no-system
类型: custommistralv3tekken
- 路径: anthracite-org/nopm_claude_writing_fixed
类型: custommistralv3tekken
- 路径: anthracite-org/kalo_opus_misc_240827_no_system
类型: custommistralv3tekken
- 路径: anthracite-org/kalo_misc_part2_no_system
类型: custommistralv3tekken
合并数据集随机打乱: true
数据集准备路径: /workspace/data/magnum-12b-data
验证集比例: 0.0
输出目录: /workspace/data/12b-fft-out
序列长度: 32768
样本打包: true
填充至序列长度: true
适配器:
LoRA模型目录:
LoRA秩:
LoRA_alpha:
LoRA丢弃率:
LoRA目标线性层:
LoRA扇入扇出:
WandB项目: 12b-magnum-fft
WandB实体:
WandB监控:
WandB名称: v4-r2-attempt-01
WandB记录模型:
梯度累积步数: 2
微批次大小: 1
训练轮数: 2
优化器: adamw_bnb_8bit
学习率调度器: cosine
学习率: 0.00001
训练输入: false
按长度分组: false
bf16: auto
fp16:
tf32: false
梯度检查点: true
早停耐心:
从检查点恢复:
本地排名:
日志步长: 1
xformers注意力:
闪光注意力: true
预热步数: 40
每轮评估次数:
评估表大小:
评估最大新标记数:
每轮保存次数: 2
调试:
深度速度: deepspeed_configs/zero2.json
权重衰减: 0.1
fsdp:
fsdp配置:
特殊标记:
填充标记: <pad>
致谢
我们要感谢Recursal / Featherless赞助本次训练的计算资源。Featherless自首个72B版本以来一直托管我们的Magnum模型,为数以千计的用户提供了访问权限,并助力我们成长。
同时感谢Anthracite全体成员使这次微调成为可能。
数据集
训练
训练进行了2轮。我们使用了由Recursal AI / Featherless AI慷慨提供的8块H100 GPU进行全参数微调。

安全性
...