49 lines
1.2 KiB
Markdown
49 lines
1.2 KiB
Markdown
|
---
|
||
|
|
library_name: transformers
|
||
|
|
license: other
|
||
|
|
tags:
|
||
|
|
- llama-factory
|
||
|
|
- full
|
||
|
|
- generated_from_trainer
|
||
|
|
model-index:
|
||
|
|
- name: onereason_full_wuliao
|
||
|
|
results: []
|
||
|
|
---
|
||
|
|
|
||
|
|
# onereason_full_wuliao
|
||
|
|
|
||
|
|
This model is a full-parameter fine-tuned OneReason-0.8B checkpoint trained on the `full_add_wuliao` dataset. This repository publishes checkpoint 600.
|
||
|
|
|
||
|
|
## Model description
|
||
|
|
|
||
|
|
The model is based on the Qwen3 causal language model architecture and is trained for OneReason multi-domain recommendation and text-to-SID tasks.
|
||
|
|
|
||
|
|
## Intended uses & limitations
|
||
|
|
|
||
|
|
This checkpoint is intended for research and evaluation on OneReason-format recommendation and SID generation tasks.
|
||
|
|
|
||
|
|
## Training data
|
||
|
|
|
||
|
|
The training dataset combines multi-domain user-history recommendation examples with material text-to-SID examples for product, live, advertisement, and video domains.
|
||
|
|
|
||
|
|
## Training procedure
|
||
|
|
|
||
|
|
Key training hyperparameters:
|
||
|
|
|
||
|
|
- learning rate: 1e-5
|
||
|
|
- train batch size per device: 1
|
||
|
|
- eval batch size per device: 1
|
||
|
|
- gradient accumulation steps: 8
|
||
|
|
- number of devices: 4
|
||
|
|
- total train batch size: 32
|
||
|
|
- seed: 42
|
||
|
|
- scheduler: linear
|
||
|
|
- warmup ratio: 0.05
|
||
|
|
- configured epochs: 3
|
||
|
|
|
||
|
|
## Framework versions
|
||
|
|
|
||
|
|
- Transformers 5.3.0
|
||
|
|
- PyTorch 2.4.1
|
||
|
|
- Tokenizers 0.22.2
|