LoRA adapter for extracting structured funding metadata (funder names + award IDs) from academic paper funding statements. Fine-tuned on Llama 3.1 8B Instruct via SFT then GRPO reinforcement learning.

Training Pipeline

Stage 1: Supervised Fine-Tuning (SFT)

Base model: meta-llama/Llama-3.1-8B-Instruct
Data: 5,264 real + 10,124 synthetic funding statements with gold-standard funder/award labels
Data augmentation: 50% of training examples augmented with synthetic noise (OCR-like case errors, digit/letter swaps, Unicode artifacts, XML/HTML tags, LaTeX markup) for robustness to real-world document formats
LoRA rank: 128
Epochs: 2

Stage 2: Reinforcement Learning (GRPO)

Algorithm: Group Relative Policy Optimization (GRPO) with importance sampling loss
Reward: Hierarchical F0.5 scoring with binary funder/award-ID matching + flat award-ID association bonus
- reward = 0.50 * funder_F0.5 + 0.50 * hierarchical_award_id_F0.5 + 0.10 * flat_award_id_F0.5
- Funder matching - Use fuzzy (token_sort_ratio ≥ 0.80 threshold, Hungarian optimal assignment)
- Award ID matching - Use binary exact after normalization (strip whitespace/hyphens/slashes, uppercase)
- Flat award-ID term - Awards partial credit when the correct award ID is extracted under the wrong funder, providing gradient on funder-award association errors
KL penalty: 0.03 (anchored to SFT checkpoint)
Group size: 8 rollouts per prompt
Temperature: 0.8
Learning rate: 3e-5
Steps: 193 batches
Checkpoint: final (batch 193)

Evaluation Results

arxiv_test.jsonl (300 held-out examples)

Permissive (partial_ratio + token_set, no damping)

Field	P	R	F1	F0.5	F1.5
Funder	0.9365	0.9408	0.9386	0.9374	0.9395
Award ID	0.9129	0.8968	0.9048	0.9096	0.9017
Scheme	0.6414	0.7686	0.6992	0.6633	0.7244
Title	0.6774	0.4375	0.5316	0.6105	0.4910

Balanced (length-damped + acronym detection)

Field	P	R	F1	F0.5	F1.5
Funder	0.8932	0.9071	0.9001	0.8960	0.9028
Award ID	0.8859	0.8702	0.8780	0.8827	0.8750
Scheme	0.5931	0.7107	0.6466	0.6134	0.6699
Title	0.6774	0.4375	0.5316	0.6105	0.4910

Strict (token_sort_ratio only)

Field	P	R	F1	F0.5	F1.5
Funder	0.8826	0.8962	0.8894	0.8853	0.8920
Award ID	0.8829	0.8673	0.8750	0.8797	0.8720
Scheme	0.5448	0.6529	0.5940	0.5635	0.6153
Title	0.6129	0.3958	0.4810	0.5523	0.4442

synthetic_edges test (1288 augmented/adversarial examples)

Permissive

Field	P	R	F1	F0.5	F1.5
Funder	0.9184	0.9343	0.9263	0.9215	0.9293
Award ID	0.8556	0.8643	0.8600	0.8573	0.8616
Scheme	0.7271	0.6774	0.7014	0.7166	0.6919
Title	0.7066	0.3430	0.4618	0.5830	0.4075

Balanced

Field	P	R	F1	F0.5	F1.5
Funder	0.8922	0.9079	0.8999	0.8953	0.9030
Award ID	0.8434	0.8520	0.8477	0.8451	0.8493
Scheme	0.6604	0.6152	0.6370	0.6509	0.6285
Title	0.6287	0.3052	0.4110	0.5188	0.3626

Strict

Field	P	R	F1	F0.5	F1.5
Funder	0.8746	0.8895	0.8820	0.8776	0.8849
Award ID	0.8366	0.8452	0.8409	0.8383	0.8425
Scheme	0.6072	0.5656	0.5857	0.5984	0.5778
Title	0.6108	0.2965	0.3992	0.5040	0.3523

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
model = PeftModel.from_pretrained(base_model, "adambuttrick/funding-extraction-llama-3.1-8b-instruct-grpo-stepfinal")
tokenizer = AutoTokenizer.from_pretrained("adambuttrick/funding-extraction-llama-3.1-8b-instruct-grpo-stepfinal")

prompt = """Extract funding information from the following statement:

This work was supported by the National Science Foundation under grant DMS-1613002 and by the NIH (R01-AI123456)."""

messages = [
    {"role": "system", "content": "You are an expert at extracting structured funding metadata from academic papers. Given a funding statement, extract all funders and their associated awards. Return a JSON array of funder objects. Each funder has:\n- \"funder_name\": string or null\n- \"awards\": array of objects with \"award_ids\" (array of strings), \"funding_scheme\" (array of strings), and \"award_title\" (array of strings)\nReturn ONLY the JSON array, no other text."},
    {"role": "user", "content": prompt},
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.0, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Output Format

json

[
  {
    "funder_name": "National Science Foundation",
    "awards": [
      {
        "award_ids": ["DMS-1613002"],
        "funding_scheme": [],
        "award_title": []
      }
    ]
  },
  {
    "funder_name": "NIH",
    "awards": [
      {
        "award_ids": ["R01-AI123456"],
        "funding_scheme": [],
        "award_title": []
      }
    ]
  }
]

cometadata/funding-extraction-llama-3.1-8b-instruct-artifact-data-mix-grpo-mixed-reward

Funding Extraction LoRA