37 lines
968 B
Markdown
37 lines
968 B
Markdown
|
---
|
||
|
|
language:
|
||
|
|
- id
|
||
|
|
license: apache-2.0
|
||
|
|
library_name: transformers
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
base_model: Qwen/Qwen2-1.5B
|
||
|
|
tags:
|
||
|
|
- indonesian
|
||
|
|
- aksarallm
|
||
|
|
- qwen2
|
||
|
|
- continued-pretrain
|
||
|
|
---
|
||
|
|
# aksarallm-1.5b-v2-checkpoint
|
||
|
|
|
||
|
|
Earlier 1.5B Qwen2-based Indonesian checkpoint. Superseded by
|
||
|
|
[`AksaraLLM/AksaraLLM-Qwen-1.5B-v5-public`](https://huggingface.co/AksaraLLM/AksaraLLM-Qwen-1.5B-v5-public)
|
||
|
|
which has measurably better perplexity and lower English leak.
|
||
|
|
|
||
|
|
## Measured baseline (Devin audit, CPU bf16, 50 short Indonesian sentences)
|
||
|
|
|
||
|
|
| Metric | Value |
|
||
|
|
|---|---|
|
||
|
|
| Perplexity | **9.9** (vs 8.4 for v5-public) |
|
||
|
|
| English-stopword ratio in ID-prompted output | 3.6% (vs 0.9% for v5-public) |
|
||
|
|
| Indonesian-stopword ratio in ID-prompted output | 26.8% |
|
||
|
|
| Parameters | 1777.1 M |
|
||
|
|
| Architecture | Qwen2ForCausalLM |
|
||
|
|
|
||
|
|
## Status
|
||
|
|
|
||
|
|
Kept for historical reference / reproducibility. **For new downstream work,
|
||
|
|
use `AksaraLLM-Qwen-1.5B-v5-public` instead.**
|
||
|
|
|
||
|
|
## License
|
||
|
|
Apache 2.0
|