Mistral-7B-Instruct-v0.3-EOSC

Federated fine tuned version using data from the EOSC registry.

Federated training configuration:

• model.name = "mistralai/Mistral-7B-Instruct-v0.3"
• model.quantization = 4
• model.gradient-checkpointing = true
• model.lora.peft-lora-r = 32
• model.lora.peft-lora-alpha = 64
• train.save-every-round = 5
• train.learning-rate-max = 5e-5
• train.learning-rate-min = 1e-6
• train.seq-length = 512
• train.training-arguments.per-device-train-batch-size = 16
• train.training-arguments.gradient-accumulation-steps = 1
• train.training-arguments.logging-steps = 10
• train.training-arguments.num-train-epochs = 2
• train.training-arguments.max-steps = 10
• train.training-arguments.save-steps = 1000
• train.training-arguments.save-total-limit = 10
• train.training-arguments.gradient-checkpointing = true
• train.training-arguments.lr-scheduler-type = "constant"
• strategy.fraction-fit = 0.1
• strategy.fraction-evaluate = 0.0
• num-server-rounds = 10

The PEFT presented in this model corresponds to 5 rounds of the FL training,

The following bitsandbytes quantization config was used during training:

• quant_method: QuantizationMethod.BITS_AND_BYTES
• _load_in_8bit: False
• _load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: fp4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float32
• bnb_4bit_quant_storage: uint8
• load_in_4bit: True
• load_in_8bit: False

Framework versions

• PEFT 0.6.2

Try the model!

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model = "mistralai/Mistral-7B-Instruct-v0.3"
adapter_model = 'ifca-advanced-computing/Mistral-7B-Instruct-v0.3-EOSC'

model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter_model)
tokenizer = AutoTokenizer.from_pretrained(base_model)

model.eval()

query = [
    {"role": "user", "content": "What is the EOSC?"},
]

input_ids = tokenizer.apply_chat_template(
    query,
    tokenize=True,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        input_ids=input_ids,
        max_new_tokens=500,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )

question = query[0]['content']
print(f'QUESTION: {question} \n')

print('ANSWER:\n')
print(tokenizer.decode(outputs[0], skip_special_tokens=True))