README.md

---
license: apache-2.0
datasets:
- ecastera/wiki_fisica
- ecastera/filosofia-es
- somosnlp/somos-clean-alpaca-es
language:
- es
- en
tags:
- mistral
- ehartford/dolphin
---

# eva-mistral-dolphin-7b-spanish

Mistral 7b based model fine tuned in Spanish to add high quality Spanish text generation.

* Base model Mistral
* Based on the excelent job of Eric Hartfod's dolphin models cognitivecomputations/dolphin-2.1-mistral-7b
* Fine-tuned in Spanish with a collection of poetry, books, wikipedia articles, phylosophy texts and alpaca-es datasets.
* Trained using Lora and PEFT on 2 GPUs for several days. 


## Usage:

Strongly advice to run inference in INT8 or INT4 mode, with the help of BitsandBytes library.


```
import torch
from transformers import AutoTokenizer, pipeline, AutoModel, AutoModelForCausalLM, BitsAndBytesConfig

MODEL = "ecastera/eva-mistral-dolphin-7b-spanish"

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    load_in_8bit=False,
    llm_int8_threshold=6.0,
    llm_int8_has_fp16_weight=False,
    bnb_4bit_compute_dtype="float16",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4")

model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    load_in_8bit=True,
    low_cpu_mem_usage=True,
    torch_dtype=torch.float16,
    quantization_config=quantization_config,
    offload_state_dict=True,
    offload_folder="./offload",
    trust_remote_code=True,
)

tokenizer = AutoTokenizer.from_pretrained(MODEL)
print(f"Loading complete {model} {tokenizer}")

prompt = "Soy Eva una inteligencia artificial y pienso que la esperanza "

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, do_sample=True, temperature=0.4, top_p=1.0, top_k=50,
                             no_repeat_ngram_size=3, max_new_tokens=100, pad_token_id=tokenizer.eos_token_id)
text_out = tokenizer.batch_decode(outputs, skip_special_tokens=True)

print(text_out)
```