Downloads · 30 days
5
11% of all-time downloads
OneReason-0.8B-web-demo11-lora
nov3630/OneReason-0.8B-web-demo11-lora
OneReason-0.8B-web-demo11-lora is a machine learning model from nov3630. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as other.
本地逐项复现网页端训练任务 demo11 (train-task-msakvu-1784869922) 的 LoRA adapter。
All-time downloads
47
Public
Repo size
56.4 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors40.4 MB · 71%
At a glance
- Library
- peft
- License
- other
- Access
- Public
- Created
- Jul 25, 2026
- Updated
- Jul 25, 2026
- SHA
- dd397bd7
Base models
Model card
From the Hugging Face model README
OneReason-0.8B web-demo11 LoRA (local repro)
本地逐项复现网页端训练任务 demo11 (train-task-msakvu-1784869922) 的 LoRA adapter。
- Base: OneReason-0.8B-pretrain-competition（快手探索者赛专用）
- 方法: LoRA r=16 α=32 dropout=0.1 target=all，SFT 3 epoch
- 超参: lr 1e-4 / linear / warmup 0.03 / wd 0.01 / bs1×accum8 / bf16 / packing
- 自定义 loss: focal γ=2 + item token ×6 加权（见
submission.py，与网页端逐字一致） - 数据: 官方三块种子数据 9 个子集，合计 32480 条
结果
| 值 | |
|---|---|
| train_loss (custom focal+item) | 1.447 |
| eval_loss (末轮) | 1.334 |
| custom_item_loss | 0.859 |
| custom_text_loss | 1.436 |
eval_loss 3 个 epoch 单调下降 (1.580→1.451→1.392→1.358→1.338→1.334)，无过拟合。
与网页端的偏离
序列长度 32768→8192（3090 显存所限，仅截断 0.12% 样本）；卡数本地 2 卡（全局 batch 16）。
完整说明见 REPRO.md。
用法
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("OpenOneRec/OneReason-0.8B-pretrain-competition", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "nov3630/OneReason-0.8B-web-demo11-lora")
- Library
- peft
- License
- other
- Created
- Jul 25, 2026
- Updated
- Jul 25, 2026