sulcantonin / llm_neuralips23

PACuna : Automated Fine-Tuning of Language Models for Particle Accelerators

- We are working on PACuna 2.0, better, more factual and public, release date 1.5.2024

Sources are in sources folder. Unforutnately, some training sources (books) are licensed, therefore neither source data nor model can be published.

pdf arxiv

.

Navigating the landscape of particle accelerators has become increasingly challenging with recent surges in contributions. These intricate devices challenge comprehension, even within individual facilities. To address this, we introduce PACuna, a fine-tuned language model refined through publicly available accelerator resources like conferences, pre-prints, and books. We automated data collection and question generation to minimize expert involvement and make the data publicly available. PACuna demonstrates proficiency in addressing intricate accelerator questions, validated by experts. Our approach shows adapting language models to scientific domains by fine-tuning technical texts and auto-generated corpora capturing the latest developments can further produce pre-trained models to answer some intricate questions that commercially available assistants cannot and can serve as intelligent assistants for individual facilities.

Code

Each LLM need a proper prompt to know that we are asking a question:

def prompt_formatter(question, answer = ""):
    return f'### Human:\n{question}\n### Assistant:\n{answer}'

PDF to Markdown

To transform PDF to MMD, we use nougat OCR (https://github.com/facebookresearch/nougat)

nougat *.pdf -o .

In publication we mistakenly omitted --no-skipping, we suggest others to use this parameter

Preprocessing

Considering Nougat OCR MultipleMarkdown, which expresses equations with (), [] and tables, we prefer to have them in standardized format where equations are expressed in $, $$ and tables in plaintext

    # replace \[.*\] with equation mode $$
    eq_expr = re.compile(r'\\\[(.*?)\\\]', flags = re.DOTALL)
    while len(re.findall(eq_expr, x)) != 0:
        eq_re = re.search(eq_expr,x)
        x = x[:eq_re.start()] + '$$' + x[eq_re.start() + 2 :eq_re.end()-2] + '$$' + x[eq_re.end():]

    # replace inline math \(.*\) with $ $
    eq_expr = re.compile(r'\\\((.*?)\\\)', flags = re.DOTALL)
    while len(re.findall(eq_expr, x)) != 0:
        eq_re = re.search(eq_expr,x)
        x = x[:eq_re.start()] + '$' + x[eq_re.start() + 2 :eq_re.end()-2] + '$' + x[eq_re.end():]
        
    # replace tables
    tabular_expr = tag_expr_full('table')
    while len(re.findall(tabular_expr,x)):
        tabular_re = re.search(tabular_expr,x)
        try:
            tabular_plain = pypandoc.convert_text(tabular_re.group(), 'markdown',format = 'latex', extra_args=['--wrap=none'])
        except Exception as e:
            tabular_plain = ""
            # print(f'EXCEPTION {str(e)} {tabular_re.group()}')
        x = x.replace(tabular_re.group(), tabular_plain)

At this stage, we have unsupervised data prepared

Supservised Data (Q&A Paris)

To generate Q&A pairs, we ask vicuna1.5-16k model with a following prompt

query = prompt_formatter(f"Generate ten questions with answers for paper:\"{data_}\"")

where data_ contains paper itself.

To generate text from LLM, we use following text block (maxl

def test(q, model, max_length=16384, temperature=0.6):
    inputs = tokenizer(q, return_tensors="pt", return_token_type_ids=False).to(device)
    outputs = model.generate(
        **inputs,
        max_length=max_length,
        # repetition_penalty=1.18,
        eos_token_id=tokenizer.eos_token_id,
        do_sample=True,
        top_p=1.0,  # top_k=64,
        temperature=temperature,
    )
    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # print(f">{answer}\n---------")
    return answer

In order to extract Q&A invidual pairs, we use following heuristics

def extract_qa_pairs_from_result(text):
    qa_pairs = []
    text = re.split(r'\d+\.\s(.*\?)\n(.*)',text)
    text = [t for t in text if not t.isspace() and len(t) > 0]
    if len(text) % 2 == 0:
        for i in range(0,len(text),2):
            q = text[i]
            a = re.sub(r'^Answer:', '', text[i+1])
            qa_pairs.append({'question' :  q,
                           'answer' : a})
    return qa_pairs

Fine-tuning LLM

See lora_unsup.ipynb

Test

    peft_model_id = model_folder + "checkpoint-x/" # Last Checkpoint Path
    
    # peft_model_id = model_folder + "/checkpoint-13200/"
    config = PeftConfig.from_pretrained(peft_model_id)
    model = AutoModelForCausalLM.from_pretrained(
        config.base_model_name_or_path,
        torch_dtype="auto",
        device_map="auto",
        offload_folder="offload",
        offload_state_dict=True,
    )  
    tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
    model = PeftModel.from_pretrained(model, peft_model_id)
    # model = model.merge_and_unload()

Querying LLM

device = 'cuda'
max_length = 256
T = 0.7

p = "What is klystron"
test(prompt_formatter(p), model, tokenizer, max_length = max_length, temperature = T)