Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
datasets:
|
| 4 |
+
- WizardLM/WizardLM_evol_instruct_V2_196k
|
| 5 |
+
- leemeng/ShareGPT90K_ja_1392
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
library_name: transformers
|
| 9 |
+
pipeline_tag: text-generation
|
| 10 |
+
tags:
|
| 11 |
+
- nlp
|
| 12 |
+
- llm
|
| 13 |
+
---
|
| 14 |
+
# AmberChat
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
We present AmberChat, an instruction following model finetuned from [LLM360/Amber](https://huggingface.co/LLM360/Amber).
|
| 18 |
+
|
| 19 |
+
## Model Description
|
| 20 |
+
|
| 21 |
+
- **Model type:** Language model with the same architecture as LLaMA-7B
|
| 22 |
+
- **Language(s) (NLP):** English
|
| 23 |
+
- **License:** Apache 2.0
|
| 24 |
+
- **Original Checkpoints:** [Aws bucket with AmberChat checkpoint with all available optimizer states](https://aws.amazon.com/)
|
| 25 |
+
- **Resources for more information:**
|
| 26 |
+
- [Research paper](https://arxiv.org/)
|
| 27 |
+
- [GitHub Repo](https://github.com/LLM360)
|
| 28 |
+
- [Amber pretraining data](https://huggingface.co/)
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
# Loading Amber
|
| 32 |
+
|
| 33 |
+
```python
|
| 34 |
+
from transformers import LlamaTokenizer, LlamaForCausalLM
|
| 35 |
+
|
| 36 |
+
tokenizer = LlamaTokenizer.from_pretrained("LLM360/AmberChat")
|
| 37 |
+
model = LlamaForCausalLM.from_pretrained("LLM360/AmberChat")
|
| 38 |
+
|
| 39 |
+
input_text = "translate English to German: How old are you?"
|
| 40 |
+
input_ids = tokenizer(input_text, return_tensors="pt").input_ids
|
| 41 |
+
|
| 42 |
+
outputs = model.generate(input_ids)
|
| 43 |
+
print(tokenizer.decode(outputs[0]))
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
# AmberChat Finetuning Details
|
| 47 |
+
|
| 48 |
+
## DataMix
|
| 49 |
+
| Subset | Number of rows |
|
| 50 |
+
| ----------- | ----------- |
|
| 51 |
+
| WizardLM/WizardLM_evol_instruct_V2_196k | 143k |
|
| 52 |
+
| Sharegpt-90k | 90k |
|
| 53 |
+
| Total | 233k |
|
| 54 |
+
|
| 55 |
+
## Hyperparameters
|
| 56 |
+
| Hyperparameter | Value |
|
| 57 |
+
| ----------- | ----------- |
|
| 58 |
+
| Total Parameters | 6.7B |
|
| 59 |
+
| Hidden Size | 4096 |
|
| 60 |
+
| Intermediate Size (MLPs) | 11008 |
|
| 61 |
+
| Number of Attention Heads | 32 |
|
| 62 |
+
| Number of Hidden Lyaers | 32 |
|
| 63 |
+
| RMSNorm ɛ | 1e^-6 |
|
| 64 |
+
| Max Seq Length | 2048 |
|
| 65 |
+
| Vocab Size | 32000 |
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
# Evaluation
|
| 69 |
+
|
| 70 |
+
| Model | MT-Bench |
|
| 71 |
+
|------------------------------------------------------|------------------------------------------------------------|
|
| 72 |
+
| LLM360/Amber 359 | 2.48750 |
|
| 73 |
+
| **LLM360/AmberChat** | **5.428125** |
|
| 74 |
+
|
| 75 |
+
# Citation
|
| 76 |
+
|
| 77 |
+
**BibTeX:**
|
| 78 |
+
|
| 79 |
+
```bibtex
|
| 80 |
+
@article{xxx,
|
| 81 |
+
title={XXX},
|
| 82 |
+
author={XXX},
|
| 83 |
+
journal={XXX},
|
| 84 |
+
year={2023}
|
| 85 |
+
}
|
| 86 |
+
```
|