Added code to use roberta-openai-detector to compare results against

Signed-off-by: Jacob Torrey <jacob@thinkst.com>

roberta_detect.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+
+# HuggingFace API test harness
+
+import re
+from typing import Optional, Tuple
+
+from roberta_local import classify_text
+
+def run_on_file_chunked(filename : str, chunk_size : int = 1024, fuzziness : int = 3) -> Optional[Tuple[str, float]]:
+	'''
+	Given a filename (and an optional chunk size) returns the score for the contents of that file.
+	This function chunks the file into at most chunk_size parts to score separately, then returns an average. This prevents a very large input
+	overwhelming the model.
+	'''
+	with open(filename, 'r') as fp:
+		contents = fp.read()
+
+	# Remove extra spaces and duplicate newlines.
+	contents = re.sub(' +', ' ', contents)
+	contents = re.sub('\t', '', contents)
+	contents = re.sub('\n+', '\n', contents)
+	contents = re.sub('\n ', '\n', contents)
+
+	start = 0
+	end = 0
+	chunks = []
+	while start + chunk_size < len(contents) and end != -1:
+		end = contents.rfind(' ', start, start + chunk_size + 1)
+		chunks.append(contents[start:end])
+		start = end + 1
+	chunks.append(contents[start:])
+	scores = []
+	for c in chunks:
+		scores.append(classify_text(c))
+	ssum : float = 0.0
+	for s in scores:
+		if s[0] == 'AI':
+			ssum -= s[1]
+		else:
+			ssum += s[1]
+	sa : float = ssum / len(scores)
+	if sa < 0:
+		return ('AI', abs(sa))
+	else:
+		return ('Human', abs(sa))
+

roberta_local.py

@@ -0,0 +1,17 @@
+#!/usr/bin/env python3
+
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("roberta-base-openai-detector")
+model = AutoModelForSequenceClassification.from_pretrained("roberta-base-openai-detector")
+
+def classify_text(s : str):
+    inputs = tokenizer(s, return_tensors='pt')
+    with torch.no_grad():
+        logits = model(**inputs).logits
+    pc = model.config.id2label[logits.argmax().item()]
+    conf = max(torch.softmax(logits, dim=1).tolist()[0])
+    if pc == 'Real':
+        return ('Human', conf)
+    return ('AI', conf)

test_roberta_detect.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+
+import pytest, os
+from warnings import warn
+from roberta_detect import run_on_file_chunked
+
+AI_SAMPLE_DIR = 'samples/llm-generated/'
+HUMAN_SAMPLE_DIR = 'samples/human-generated/'
+
+ai_files = os.listdir(AI_SAMPLE_DIR)
+human_files = os.listdir(HUMAN_SAMPLE_DIR)
+
+CONFIDENCE_THRESHOLD : float = 0.00 # What confidence to treat as error vs warning
+
+def test_training_file():
+    (classification, score) = run_on_file_chunked('ai-generated.txt')
+    assert classification == 'AI', 'The training corpus should always be detected as AI-generated... since it is (score: ' + str(round(score, 8)) + ')'
+
+@pytest.mark.parametrize('f', human_files)
+def test_human_samples(f):
+    (classification, score) = run_on_file_chunked(HUMAN_SAMPLE_DIR + f)
+    if score > CONFIDENCE_THRESHOLD:
+        assert classification == 'Human', f + ' is a human-generated file, misclassified as AI-generated with confidence ' + str(round(score, 8))
+    else:
+        if classification != 'Human':
+            warn("Misclassified " + f + " with score of: " + str(round(score, 8)))
+        else:
+            warn("Unable to confidently classify: " + f)
+
+@pytest.mark.parametrize('f', ai_files)
+def test_llm_sample(f):
+   (classification, score) = run_on_file_chunked(AI_SAMPLE_DIR + f)
+   if score > CONFIDENCE_THRESHOLD:
+       assert classification == 'AI', f + ' is an LLM-generated file, misclassified as human-generated with confidence ' + str(round(score, 8))
+   else:
+       if classification != 'AI':
+           warn("Misclassified " + f + " with score of: " + str(round(score, 8)))
+       else:
+           warn("Unable to confidently classify: " + f)