license: apache-2.0
language:
- zh
- en
Llama-3-Chinese-8B-GGUF
Warning: llama.cpp has breaking changes on Llama-3 pre-tokenizer, which significantly affect performance. We will update GGUF mdoels in the next few hours.
This repository contains Llama-3-Chinese-8B-GGUF (llama.cpp/ollama/tgw, etc. compatible), which is the quantized version of Llama-3-Chinese-8B.
Note: this is a foundation model, which is not suitable for conversation, QA, etc.
Further details (performance, usage, etc.) should refer to GitHub project page: https://github.com/ymcui/Chinese-LLaMA-Alpaca-3
Performance
Metric: PPL, lower is better
The model name with -im suffix is generated with important matrix, which has generally better performance.
| Quant | Size | PPL | PPL (-im) |
|---|---|---|---|
| Q2_K | 2.96 GB | 17.7212 +/- 0.59814 | 14.9583 +/- 0.50455 |
| Q3_K | 3.74 GB | 8.6303 +/- 0.28481 | 8.4423 +/- 0.28087 |
| Q4_0 | 4.34 GB | 8.2513 +/- 0.27102 | 7.9077 +/- 0.25525 |
| Q4_K | 4.58 GB | 7.8897 +/- 0.25830 | 7.8279 +/- 0.25542 |
| Q5_0 | 5.21 GB | 7.7975 +/- 0.25639 | 7.7724 +/- 0.25625 |
| Q5_K | 5.34 GB | 7.7062 +/- 0.25218 | 7.6902 +/- 0.25170 |
| Q6_K | 6.14 GB | 7.6600 +/- 0.25043 | 7.6412 +/- 0.24949 |
| Q8_0 | 7.95 GB | 7.6512 +/- 0.25064 | 7.6512 +/- 0.25064 |
| F16 | 14.97 GB | 7.6389 +/- 0.25001 | N/A |
Others
For full model, please see: https://huggingface.co./hfl/llama-3-chinese-8b
For LoRA-only model, please see: https://huggingface.co./hfl/llama-3-chinese-8b-lora
If you have questions/issues regarding this model, please submit an issue through https://github.com/ymcui/Chinese-LLaMA-Alpaca-3