Lux-In-Tenebris/research_summary_deconstructor

Fine-tuned version of IBM Granite 3.3 2b base

Accepts a research project title and summary and creates a largely extractive breakdown (with some minor tweaks for coherence) into:

• context and background
• problem and aim
• approach and methodology
• outcomes and impact

with seperate keywords for each of these breakdowns.

Additionally includes lists of:

• application areas - where the research is indicated as applying to
• expected beneficiaries - where indicated within the document

Usage:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import json

input_title: str = "your project title here"
input_summary: str = "your summary here"
model_name: str = "Lux-In-Tenebris/research_summary_deconstructor"

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForCausalLM.from_pretrained(model_name, dtype=torch.bfloat16).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model.eval()

input_text = json.dumps({"title": input_title, "summary": input_summary}, ensure_ascii=False)
inputs = tokenizer(input_text, return_tensors="pt", truncation=False, padding=False).to(device)

with torch.no_grad():
  outputs = model.generate(**inputs, max_new_tokens=8_192, temperature=1.0, do_sample=True)
generated_tokens = outputs[0][inputs["input_ids"].shape[1]:]
decoded = tokenizer.decode(generated_tokens, skip_special_tokens=True)

Will of course need adjustment for batch processing, and highly recommend checking for invalid json and additional characters being output.

Trained on approx 5000 synthetic research project titles, summaries, and structured elements, for 2 epochs using an RTX6000 Pro

• Learning rate: 3e-5
• Effective batch size: 128
• Time taken: Under 2 hours