Model Card: Clinical Trial Power Analysis — Llama 3.3 70B LoRA

Model Details

Model Description

This is a LoRA adapter fine-tuned on top of Llama-3.3-70B-Instruct to perform statistical power classification for clinical and research trial designs. Given a trial's design — test type, sample size, expected effect size, and significance level — the model computes the actual statistical power achieved, classifies the trial as adequately powered or underpowered, and states the minimum sample size that would be needed to reach the conventional 80% power threshold.

This targets a well-documented, real problem in scientific credibility: a substantial share of irreproducible findings across psychology, medicine, and other fields trace back to studies that were underpowered from the start — enrolled too few participants to reliably detect the effect they set out to test. Correctly classifying power requires genuine computation (the noncentral-distribution relationship between sample size, effect size, and detectable power), not verbal pattern-matching, which is a documented weakness in how base LLMs reason about statistical adequacy.

Training labels were generated by construction from the standard, textbook two-sample power approximation formula (n_per_group = 2 × ((z_α/2 + z_β) / d)², the widely-taught closed-form approximation for two-sided tests at 80% power), so the classification, required sample size, and enrollment gap are all guaranteed internally consistent and independently re-verifiable from the reported numbers — not separately annotated. This formula was additionally sanity-checked against statsmodels.stats.power (the standard, peer-vetted Python implementation used in real biostatistics work) before dataset construction.

Submitted to the AutoScientist Challenge (Science track) by Adaption Labs.

Developed by: Grold Otieno Mboya
Shared by: Grold Otieno Mboya
Model type: LoRA adapter for causal language modeling (statistical power / trial design reasoning)
Language(s): English
License: Llama 3.3 Community License Agreement (inherited from the base model — see https://www.llama.com/llama3_3/license/)
Finetuned from model: meta-llama/Llama-3.3-70B-Instruct

Model Sources

Platform	Dataset	Model
Hugging Face	adaption-clinical-trial-power-analysis	clinical-trial-power-analysis-llama3-3-70b-lora
Kaggle	adaption-clinical-trial-power-analysis	clinical-trial-power-analysis-llama3-3-70b-lora

Live Demo: trialpower.adaptionlabs.app

Uses

Direct Use

Checking whether a planned or completed trial's sample size is actually sufficient to detect its claimed effect size: useful for grant reviewers and funding committees screening proposals, IRB/ethics committees assessing trial design, journal reviewers evaluating submitted studies, and early-career researchers or PhD candidates building intuition around sample-size decisions before designing a study.

Downstream Use

Could be extended to additional test types beyond the current coverage (e.g. chi-square tests, survival/Cox regression, cluster-randomized or non-inferiority designs), or integrated into grant-submission or IRB-review tooling as an automated pre-screening step.

Out-of-Scope Use

Not a substitute for a full consultation with a professional biostatistician, particularly for complex designs (multi-arm, adaptive, cluster-randomized, or non-inferiority trials) that the closed-form approximation used here does not model. Not evaluated for general capability, factuality, or safety beyond the Science task category described here. The underlying formula is a standard approximation, not the exact noncentral-t solution — adequate for the two-sample/paired/one-sample designs it targets, but not a replacement for exact power software on borderline or unusual cases.

Bias, Risks, and Limitations

The training data combines 15 real, cited rows sourced from actual ClinicalTrials.gov filings (spot-checked for authenticity — e.g. NCT06983743, a real Phase 1 trial of ERAS-0015 in advanced solid tumors, confirmed against independent sources) with 3,100 synthetic rows using invented trial/company names but realistic parameter ranges. All labels — required sample size, enrollment gap, and power classification — were independently re-derived from the raw numbers using the stated formula and confirmed to match exactly across all 3,115 rows, with zero inconsistencies found.

Known class imbalance: the dataset skews toward adequately-powered examples (2,134 adequately powered vs. 981 underpowered, roughly a 2:1 split) rather than an even balance — worth accounting for when interpreting edge-case performance.

Known approximation limitation: the closed-form formula used for ground truth is the standard normal-approximation, not the exact noncentral-t calculation; for small sample sizes or extreme effect sizes, professional power-analysis software may compute a slightly different value than this model reports.

Known validation gap: authenticity of the 15 real anchor rows was spot-checked for trial identity and design details, but exact enrollment figures were not independently re-confirmed against the live registry for every row. Performance has only been measured against Adaption Labs' own evaluation sets (see Results) — not yet tested against a larger, independent sample of real trial registry data.

Recommendations

Treat outputs as a rapid pre-screening signal, not a final determination — borderline classifications, complex trial designs, or high-stakes funding/publication decisions should be confirmed with exact power-analysis software or a professional biostatistician.

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "meta-llama/Llama-3.3-70B-Instruct"
adapter_id = "Gro97/clinical-trial-power-analysis-llama3-3-70b-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_id)

# NOTE: reconstructed to match the verified dataset schema and formula --
# check against the actual "system" field in your training file before
# relying on this verbatim, since it was not authored in this session.
system_prompt = (
    "You are an expert biostatistician evaluating clinical trial designs. "
    "Given a trial's test type, sample size, expected effect size, and alpha "
    "level, compute the statistical power achieved, classify the trial as "
    "adequately_powered or underpowered (80% power is the standard threshold), "
    "and state the minimum sample size required for 80% power. Respond only "
    "with strict JSON with exactly these keys: trial_id, condition, phase, "
    "actual_enrollment, assumed_effect_size_label, assumed_cohens_d, "
    "required_n_for_80_percent_power, power_classification, enrollment_gap, "
    "interpretation."
)
user_prompt = (
    "Trial NCT-EXAMPLE-001, Phase 2, studying a novel treatment for a "
    "chronic condition. Actual enrollment: 120 participants. Assumed "
    "effect size: medium (Cohen's d = 0.5)."
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_prompt},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=400)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

Training Data

Custom-built dataset clinical_trial_power_analysis: 3,115 rows.

Component	Rows	Description
Real (cited)	15	Sourced from actual ClinicalTrials.gov filings, spot-checked for authenticity
Synthetic (realistic)	3,100	Invented trial identifiers, realistic parameter ranges

Phase coverage spans Phase 1, Phase 1/Phase 2, Phase 2, Phase 3, Phase 4, and a small number of observational designs. Effect sizes are roughly evenly split across small/medium/large (Cohen's d conventions). Power classification skews toward adequately-powered (2,134) over underpowered (981) — see Limitations above.

Every row's required_n_for_80_percent_power, enrollment_gap, and power_classification were independently recomputed from assumed_cohens_d and actual_enrollment using the stated formula and confirmed to match exactly, with zero inconsistencies across all 3,115 rows.

Training Procedure

Fine-tuned via Adaption Labs' AutoScientist platform using LoRA.

Training Hyperparameters

Training regime: see adapter_config.json in this repository for the exact LoRA rank/alpha/target-module configuration used for this run.

Speeds, Sizes, Times

Not independently benchmarked by the author; training was run on Adaption Labs' hosted infrastructure (specific hardware not disclosed to the end user for this run).

Evaluation

Testing Data, Factors & Metrics

Testing Data

Two evaluation sets, as reported by Adaption Labs' AutoScientist platform: (1) a held-out split of this project's own dataset, (2) a broader set of unseen tasks from Adaption's internal Science category benchmark.

Metrics

Pairwise preference win rate against the un-adapted base model, as reported by Adaption Labs' AutoScientist platform (judge methodology not disclosed to the end user).

Results

Evaluation set	Base model win rate	Adapted model win rate
This dataset's held-out samples	23%	77%
Broader Science category	24%	76%

Summary

The adapter shows a large improvement over the base model both on its own training distribution (77% win rate) and on the broader Science category benchmark (76% win rate) — the two numbers being close suggests the gains generalize well beyond the specific training distribution, consistent with the model having learned the underlying power-computation skill rather than memorizing dataset-specific patterns. This is the strongest result across this author's AutoScientist submissions, consistent with the hypothesis that tasks requiring genuine numeric computation (rather than pattern recognition of well-known concepts) produce the largest, most defensible base-to-adapted gaps.

Environmental Impact

Hardware Type: Not disclosed by the training platform for this run.
Hours used: Not disclosed.
Cloud Provider: Adaption Labs' hosted infrastructure.
Compute Region: Not disclosed.
Carbon Emitted: Not calculated.

Technical Specifications

Model Architecture and Objective

LoRA adapter applied to meta-llama/Llama-3.3-70B-Instruct, trained via supervised fine-tuning to compute and classify statistical power for clinical/research trial designs.

Compute Infrastructure

Software

PEFT 0.15.1
Trained via Adaption Labs' AutoScientist platform

Citation

BibTeX:

@misc{mboya2026trialpower,
  author = {Mboya, Grold Otieno},
  title = {Clinical Trial Power Analysis: A LoRA Adapter for Statistical
           Power Classification of Trial Designs},
  year = {2026},
  howpublished = {AutoScientist Challenge submission, Adaption Labs},
  url = {https://huggingface.co/Gro97/clinical-trial-power-analysis-llama3-3-70b-lora}
}

Model Card Contact

See author's Hugging Face profile: https://huggingface.co/Gro97

Framework versions

PEFT 0.15.1

Downloads last month: 3

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Gro97/clinical-trial-power-analysis-llama3-3-70b-lora

Base model

meta-llama/Llama-3.1-70B

Finetuned

meta-llama/Llama-3.3-70B-Instruct

Adapter

(442)

this model