training/eval.py

from unsloth import FastLanguageModel
import torch,sys

model_name_input = sys.argv[1]

max_seq_length = 4096 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

model, tokenizer = FastLanguageModel.from_pretrained(
    #model_name = "unsloth/mistral-7b-instruct-v0.2-bnb-4bit", # Choose ANY! eg teknium/OpenHermes-2.5-Mistral-7B
    model_name = model_name_input,
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
def formatting_prompts_func(examples):
    instructions = examples["instruction"]
    inputs       = examples["input"]
    outputs      = examples["output"]
    texts = []
    for instruction, input, output in zip(instructions, inputs, outputs):
        # Must add EOS_TOKEN, otherwise your generation will go on forever!
        text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts, }
pass

#load and convert the dataset into the prompt format
from datasets import load_dataset
dataset = load_dataset("json", data_files="data.json", split = "train")
dataset = dataset.map(formatting_prompts_func, batched = True,)

FastLanguageModel.for_inference(model)
# do x evals of items from the dataset before training
samples = []
sample_size = 10
for x in range(0,sample_size):
    instruction = dataset[x]["instruction"]
    input       = dataset[x]["input"]
    output      = ''
    text = alpaca_prompt.format(instruction, input, output) #+ EOS_TOKEN
    sample = tokenizer([text],return_tensors = "pt").to("cuda")
    out = model.generate(**sample,max_new_tokens=4096,use_cache=True)
    out = tokenizer.batch_decode(out)
    samples.append(out[0])

# new one not in your dataset goes here
code = '''int __fastcall sub_75C80(int a1, int a2)
{
  int result; // r0
  _DWORD *i; // r3

  result = a2 - *(_DWORD *)(a1 + 12);
  for ( i = *(_DWORD **)(a1 + 48); i; i = (_DWORD *)*i )
  {
    if ( i[2] < result )
      result = i[2];
  }
  return result;
}'''

text = alpaca_prompt.format(instruction, code, output)
sample = tokenizer([text],return_tensors = "pt").to("cuda")
out = model.generate(**sample,max_new_tokens=4096,use_cache=True)
out = tokenizer.batch_decode(out)
samples.append(out[0])

print('Capturing pre training generation samples')
with open(f'results/eval_log_{model_name_input.replace("/","_")}','w') as log:
    for r in samples:
        log.write(r)