mcqgen_test1 / README.md
librarian-bot's picture
Librarian Bot: Add base_model information to model
a20177a
|
raw
history blame
2.74 kB
metadata
library_name: peft
base_model: meta-llama/Llama-2-7b-hf

Training procedure

The following bitsandbytes quantization config was used during training:

  • 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

Framework versions

  • PEFT 0.4.0

notebook (training and inference): https://colab.research.google.com/drive/1GxbUYZiLidteVX4qu5iSox6oxxEOHk5O?usp=sharing

Usage:

import requests

# Get a random Wikipedia article summary using their API
def random_extract():
    URL = "https://en.wikipedia.org/api/rest_v1/page/random/summary"
    PARAMS = {}
    r = requests.get(url = URL, params = PARAMS)
    data = r.json()
    return data['extract']

# Format this as a prompt that would hopefully result in the model completing with a question
def random_prompt():
  e = random_extract()
  return f"""### CONTEXT: {e} ### QUESTION:"""

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

output_dir = "mcqgen_test"

# load base LLM model and tokenizer
model = AutoPeftModelForCausalLM.from_pretrained(
    output_dir,
    low_cpu_mem_usage=True,
    torch_dtype=torch.float16,
    load_in_4bit=True,
)
tokenizer = AutoTokenizer.from_pretrained(output_dir)

# We can feed in a random context prompt and see what question the model comes up with:
prompt = random_prompt()

input_ids = tokenizer(prompt, return_tensors="pt", truncation=True).input_ids.cuda()
# with torch.inference_mode():
outputs = model.generate(input_ids=input_ids, max_new_tokens=100, do_sample=True, top_p=0.9,temperature=0.9)

print(f"Prompt:\n{prompt}\n")
print(f"Generated MCQ:\n### QUESTION:{tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0][len(prompt):]}")

def process_outputs(outputs):
  s = tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0]
  split = s.split("### ")[1:][:7]
  if len(split) != 7:
    return None
  # Check the starts
  expected_starts = ['CONTEXT', 'QUESTION', 'A' , 'B', 'C', 'D', 'CORRECT']
  for i, s in enumerate(split):
    if not split[i].startswith(expected_starts[i]):
      return None
  return {
      "context": split[0].replace("CONTEXT: ", ""),
      "question": split[1].replace("QUESTION: ", ""),
      "a": split[2].replace("A: ", ""),
      "b": split[3].replace("B: ", ""),
      "c": split[4].replace("C: ", ""),
      "d": split[5].replace("D: ", ""),
      "correct": split[6].replace("CORRECT: ", "")
  }


process_outputs(outputs) # A nice dictionary hopefully