File size: 1,555 Bytes
cfa7fff |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 |
---
license: apache-2.0
datasets:
- lodrick-the-lafted/Hermes-217K
---
<img src=https://huggingface.co/lodrick-the-lafted/Hermes-Instruct-217K/resolve/main/hermes-instruct.png>
# Hermes-Instruct-7B-217K
[Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2) trained with 217K rows of [teknium/openhermes](https://huggingface.co/datasets/teknium/openhermes), in Alpaca format.
Why? Mistral-7B-Instruct-v0.2 has native 32K context and rope theta of 1M. It's not a base model, so I've used the same recipe with different amounts of data to gauge the effects of further finetuning.
<br />
<br />
# Prompt Format
Both the default Mistral-Instruct tags and Alpaca are fine, so either:
```
<s>[INST] {sys_prompt} {instruction} [/INST]
```
or
```
{sys_prompt}
### Instruction:
{instruction}
### Response:
```
The tokenizer default is Alpaca this time around.
<br />
<br />
# Usage
```python
from transformers import AutoTokenizer
import transformers
import torch
model = "lodrick-the-lafted/Hermes-Instruct-7B-217K"
tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
"text-generation",
model=model,
model_kwargs={"torch_dtype": torch.bfloat16},
)
messages = [{"role": "user", "content": "Give me a cooking recipe for an apple pie."}]
prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.95)
print(outputs[0]["generated_text"])
```
|