---Statistical Consistency Checker

Category: Science | Model: Llama-4-Scout-17B-16E-Instruct (LoRA)

Problem

Nuijten et al. (2015) scanned over 250,000 p-values across major psychology journals and found that roughly half of all published papers using null-hypothesis significance testing contained at least one p-value inconsistent with its reported test statistic and degrees of freedom. About one in eight papers contained a "gross" inconsistency severe enough to potentially flip the paper's statistical conclusion. The existing detection tool (statcheck) is a regex-based R package restricted to strict APA-formatted statistics, unable to parse narrative-style reporting or check whether a paper's prose claims actually match its own reported numbers.

Solution

Closed-label classification (consistent / minor_inconsistency / gross_inconsistency) grounded in exact recomputed mathematics rather than inference. For each reported statistic, the true p-value is computed directly from the relevant distribution:

t-test → two-tailed p from the Student's t CDF F-test → p from the F distribution CDF Chi-square → p from the chi-square distribution CDF Correlation (r) → converted to an equivalent t statistic, then t-distribution math

Classification thresholds mirror the real, published statcheck methodology: a "gross" inconsistency is defined as the reported and recomputed p-values falling on opposite sides of alpha = 0.05 (i.e., the paper's actual conclusion would flip), not merely a numeric mismatch.

Dataset

10,000 rows generated via Python/scipy, every recomputed_p value independently reproducible by rerunning the same computation against the same reported statistic and degrees of freedom — zero LLM-generated or estimated labels anywhere in the pipeline. Balanced across all 4 test types and 3 consistency categories (~36/30/33%), with genuine prompt-level diversity (7 rotating phrasing templates, zero repeated prompt text across all 10,000 rows).

Results Metric Base Adapted On-dataset win rate 39 61 Science-category win rate 30 70 Dataset quality grade E B (+190.0%) Custom rubric score 1.6 9.1 (+461.7%) Limitations

Covers single-sentence statistical claims across 4 common test types; real papers report results in more varied, multi-sentence, and non-English forms. Per-class precision/recall (particularly gross_inconsistency recall) has not yet been measured against the independent 300-row holdout set built for this purpose — a documented next step, not a hidden gap. Label balance was deliberately set for training purposes and does not reflect real-world error prevalence (Nuijten et al. found ~10-15% of individual p-values inconsistent in actual literature).

Assets

Dataset card, model card, standalone HTML demo (real browser-computed statistical math, verified against scipy to floating-point precision) — all built and ready. base_model: meta-llama/Llama-4-Scout-17B-16E-Instruct library_name: peft license: other tags: - lora - peft - adapter - adaption

adaption_stats_consistency_labels

Model Training

A LORA adapter for meta-llama/Llama-4-Scout-17B-16E-Instruct. This model was trained with SFT using Adaption's AutoScientist on the stats_consistency_labels dataset.

AutoScientist Config

{
  "job_id": "41d24b78-580c-4bcb-be31-acdc7c957319",
  "training_experiment_id": "d6c2ed0b-4b00-4a4b-a11a-753366703955",
  "original_model_name": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
  "trained_model_name": "adaption_stats_consistency_labels",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 64,
    "n_evals": 5,
    "n_epochs": 3,
    "batch_size": "max",
    "lora_alpha": 128,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.05,
    "weight_decay": 0.03,
    "learning_rate": 0.0001,
    "max_grad_norm": 1,
    "base_model_size": "109B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "k_proj,o_proj,q_proj,v_proj,shared_expert.gate_proj,shared_expert.up_proj,shared_expert.down_proj,feed_forward.gate_proj,feed_forward.up_proj,feed_forward.down_proj"
  }
}

Training Data

The model was trained on 27,801 rows of adapted data with the following domain distribution: science (56%), data-analysis-visualization (28%), math (15%), academic-education (0%), academia-education (0%).

Model Evaluation

The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.

Domain	Win rate vs. base model
science	70%

How to use

pip install torch transformers peft

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
ADAPTER = "<this-repo-id>"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16

base = AutoModelForCausalLM.from_pretrained(BASE, dtype=dtype).to(device)
model = PeftModel.from_pretrained(base, ADAPTER)
# Optional: merge the LoRA weights into the base for faster inference
model = model.merge_and_unload()
model.eval()

tokenizer = AutoTokenizer.from_pretrained(BASE)
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(device)

with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))