jit-lora / tests /test_statistical_e2e.py
Ex0bit's picture
Upload complete JIT LoRA research: paper, source code, tests, and figures
208eb59
#!/usr/bin/env python3
"""
test_statistical_e2e.py — Statistically rigorous JIT LoRA training evaluation.
Dynamically loads real-world facts (post model training cutoff), pre-tests each
against the model to confirm it's truly unknown, trains via LoRA, then evaluates
with proper statistical analysis across multiple independent trials.
Usage:
# Ensure daemon is running with model activated
python3 test_statistical_e2e.py
# Custom options
python3 test_statistical_e2e.py --facts-file raw_facts_2026.txt --trials 3 --max-facts 80
Data source: facts are loaded from a file generated by web search (not hardcoded).
The file format is:
CATEGORY: <category>
Q: <question>
A: <answer>
KEYWORDS: <comma-separated keywords>
"""
import argparse
import json
import math
import os
import random
import re
import statistics
import sys
import time
from dataclasses import dataclass, field
from pathlib import Path
from typing import Optional
import requests
# ─── Configuration ───────────────────────────────────────────────────────────
DAEMON_URL = "http://localhost:8766"
DEFAULT_FACTS_FILE = os.path.join(os.path.dirname(__file__), "raw_facts_2026.txt")
MAX_TOKENS_PRETEST = 80 # Short response — just need to check if it knows
MAX_TOKENS_POSTTEST = 100 # Enough for a factual answer
TRAIN_EPOCHS = 15
REGULARIZATION_RATIO = 0.33 # ≥33% real-world data to prevent catastrophic forgetting
# ─── Data Structures ────────────────────────────────────────────────────────
@dataclass
class Fact:
category: str
question: str
answer: str
keywords: list # minimum keywords to verify correctness
pretest_response: str = ""
pretest_known: bool = False # True if model already knows this fact
def to_training_pair(self):
return {"user": self.question, "assistant": self.answer}
@dataclass
class TrialResult:
trial_id: int
n_facts_tested: int
n_confirmed_unknown: int
n_training_pairs: int
n_regularization_pairs: int
training_steps: int
training_time_s: float
initial_loss: float
final_loss: float
# Post-training scores (count correct)
recall_correct: int
recall_total: int
general_correct: int
general_total: int
# Per-category breakdown
category_scores: dict = field(default_factory=dict)
# ─── Fact Loading ────────────────────────────────────────────────────────────
def load_facts_from_file(filepath: str) -> list:
"""Parse the raw facts file into Fact objects."""
facts = []
current = {}
with open(filepath) as f:
for line in f:
line = line.strip()
if not line or line.startswith("#") or line.startswith("="):
continue
if line.startswith("CATEGORY:"):
if current.get("question"):
facts.append(Fact(
category=current.get("category", "Unknown"),
question=current["question"],
answer=current.get("answer", ""),
keywords=[k.strip().lower() for k in current.get("keywords", "").split(",") if k.strip()],
))
current = {"category": line.split(":", 1)[1].strip()}
elif line.startswith("Q:"):
# If we have a pending fact, save it first
if current.get("question"):
facts.append(Fact(
category=current.get("category", "Unknown"),
question=current["question"],
answer=current.get("answer", ""),
keywords=[k.strip().lower() for k in current.get("keywords", "").split(",") if k.strip()],
))
cat = current.get("category", "Unknown")
current = {"category": cat}
current["question"] = line[2:].strip()
elif line.startswith("A:"):
current["answer"] = line[2:].strip()
elif line.startswith("KEYWORDS:"):
current["keywords"] = line[9:].strip()
# Don't forget the last fact
if current.get("question"):
facts.append(Fact(
category=current.get("category", "Unknown"),
question=current["question"],
answer=current.get("answer", ""),
keywords=[k.strip().lower() for k in current.get("keywords", "").split(",") if k.strip()],
))
return facts
# ─── General Knowledge Test Set ──────────────────────────────────────────────
GENERAL_KNOWLEDGE = [
{"question": "What is the capital of France?", "keywords": ["paris"]},
{"question": "Who wrote Romeo and Juliet?", "keywords": ["shakespeare"]},
{"question": "What is the chemical symbol for water?", "keywords": ["h2o"]},
{"question": "What planet is closest to the Sun?", "keywords": ["mercury"]},
{"question": "What year did World War II end?", "keywords": ["1945"]},
{"question": "What is the speed of light in km/s approximately?", "keywords": ["299", "km"]},
{"question": "Who painted the Mona Lisa?", "keywords": ["vinci", "leonardo"]},
{"question": "What is the largest ocean on Earth?", "keywords": ["pacific"]},
{"question": "What gas do plants absorb from the atmosphere?", "keywords": ["co2", "carbon dioxide"]},
{"question": "What is the square root of 144?", "keywords": ["12"]},
{"question": "Who developed the theory of general relativity?", "keywords": ["einstein"]},
{"question": "What is the capital of Japan?", "keywords": ["tokyo"]},
{"question": "How many chromosomes do humans have?", "keywords": ["46", "23 pairs"]},
{"question": "What element has the atomic number 1?", "keywords": ["hydrogen"]},
{"question": "Who was the first person to walk on the Moon?", "keywords": ["armstrong"]},
{"question": "What is the boiling point of water in Celsius?", "keywords": ["100"]},
{"question": "What is the capital of Australia?", "keywords": ["canberra"]},
{"question": "What year was the United Nations founded?", "keywords": ["1945"]},
{"question": "What is the chemical formula for table salt?", "keywords": ["nacl"]},
{"question": "Who wrote 1984?", "keywords": ["orwell"]},
]
# ─── Regularization pairs (real-world Q&A to prevent forgetting) ─────────────
REGULARIZATION_PAIRS = [
{"user": "What is the capital of France?", "assistant": "The capital of France is Paris."},
{"user": "Who wrote Romeo and Juliet?", "assistant": "William Shakespeare wrote Romeo and Juliet."},
{"user": "What is the chemical symbol for water?", "assistant": "The chemical symbol for water is H2O."},
{"user": "What planet is closest to the Sun?", "assistant": "Mercury is the closest planet to the Sun."},
{"user": "What year did World War II end?", "assistant": "World War II ended in 1945."},
{"user": "Who painted the Mona Lisa?", "assistant": "Leonardo da Vinci painted the Mona Lisa."},
{"user": "What is the largest ocean on Earth?", "assistant": "The Pacific Ocean is the largest ocean on Earth."},
{"user": "What gas do plants absorb from the atmosphere?", "assistant": "Plants absorb carbon dioxide (CO2) from the atmosphere."},
{"user": "What is the square root of 144?", "assistant": "The square root of 144 is 12."},
{"user": "Who developed the theory of general relativity?", "assistant": "Albert Einstein developed the theory of general relativity."},
{"user": "What is the capital of Japan?", "assistant": "The capital of Japan is Tokyo."},
{"user": "How many chromosomes do humans have?", "assistant": "Humans have 46 chromosomes, or 23 pairs."},
{"user": "What element has the atomic number 1?", "assistant": "Hydrogen has the atomic number 1."},
{"user": "Who was the first person to walk on the Moon?", "assistant": "Neil Armstrong was the first person to walk on the Moon in 1969."},
{"user": "What is the boiling point of water in Celsius?", "assistant": "The boiling point of water is 100 degrees Celsius."},
{"user": "What is the capital of Australia?", "assistant": "The capital of Australia is Canberra."},
{"user": "What year was the United Nations founded?", "assistant": "The United Nations was founded in 1945."},
{"user": "What is the chemical formula for table salt?", "assistant": "The chemical formula for table salt is NaCl (sodium chloride)."},
{"user": "Who wrote the novel 1984?", "assistant": "George Orwell wrote the novel 1984."},
{"user": "What is the tallest mountain in the world?", "assistant": "Mount Everest is the tallest mountain in the world at 8,849 meters."},
]
# ─── Daemon API ──────────────────────────────────────────────────────────────
def daemon_status():
r = requests.get(f"{DAEMON_URL}/status", timeout=10)
r.raise_for_status()
return r.json()
def daemon_reset(retries=3):
"""Reset adapter and data buffers for a clean trial."""
for attempt in range(retries):
try:
r = requests.post(f"{DAEMON_URL}/reset", json={"clear_data": True}, timeout=60)
r.raise_for_status()
return r.json()
except Exception as e:
if attempt < retries - 1:
print(f" Reset attempt {attempt+1} failed: {e}, retrying in 5s...")
time.sleep(5)
else:
raise
def daemon_query(question: str, max_tokens: int = 100) -> str:
"""Query the model and collect the full response."""
try:
r = requests.post(
f"{DAEMON_URL}/chat",
json={"messages": [{"role": "user", "content": question}],
"max_tokens": max_tokens, "stream": True},
stream=True, timeout=180,
)
r.raise_for_status()
except Exception as e:
print(f" [Query error: {e}]")
return ""
text = ""
try:
for line in r.iter_lines(decode_unicode=True):
if not line or not line.startswith("data: "):
continue
payload = line[6:].strip()
if payload == "[DONE]":
break
try:
obj = json.loads(payload)
delta = obj.get("choices", [{}])[0].get("delta", {})
content = delta.get("content", "")
# Filter out special tokens
if content and not content.startswith("<|"):
text += content
except json.JSONDecodeError:
continue
except Exception as e:
print(f" [Stream error: {e}, got so far: {text[:50]}]")
return text.strip()
def daemon_inject_and_train(training_pairs: list, epochs: int = TRAIN_EPOCHS) -> dict:
"""Inject training data and run epoch-based training.
Converts {"user": ..., "assistant": ...} pairs to the daemon's expected format:
[{"role": "user", "content": ...}, {"role": "assistant", "content": ...}]
The /train endpoint is async — it starts training in background and returns immediately.
We poll /status until training completes.
"""
# Convert pair format to message format
messages = []
for pair in training_pairs:
messages.append([
{"role": "user", "content": pair["user"]},
{"role": "assistant", "content": pair["assistant"]},
])
r = requests.post(
f"{DAEMON_URL}/train",
json={"messages": messages, "epochs": epochs},
timeout=30,
)
r.raise_for_status()
start_response = r.json()
print(f" Train started: injected={start_response.get('injected', 0)}, epochs={start_response.get('epochs', 0)}")
# Poll until training completes
poll_interval = 2
max_wait = 600 # 10 minutes max
elapsed = 0
last_steps = 0
result = {"steps": 0, "final_loss": 0, "initial_loss": 0, "epochs_completed": 0, "early_stopped": False}
while elapsed < max_wait:
time.sleep(poll_interval)
elapsed += poll_interval
try:
status = daemon_status()
current_steps = status.get("total_steps", 0)
current_loss = status.get("last_loss", 0)
if current_steps != last_steps:
last_steps = current_steps
if not status.get("training", False):
# Training finished
result["steps"] = status.get("total_steps", 0)
result["final_loss"] = status.get("last_loss", 0)
result["initial_loss"] = result.get("initial_loss", current_loss)
break
# Update initial loss from first poll
if result["initial_loss"] == 0 and current_loss > 0:
result["initial_loss"] = current_loss
if elapsed % 30 == 0:
print(f" ... training: step={current_steps}, loss={current_loss:.4f}")
except Exception as e:
print(f" [Poll error: {e}]")
return result
def daemon_set_auto_train(enabled: bool):
"""Enable/disable auto_train on the daemon."""
try:
r = requests.put(
f"{DAEMON_URL}/config",
json={"auto_train": enabled},
timeout=10,
)
r.raise_for_status()
except Exception as e:
print(f" [Warning: could not set auto_train={enabled}: {e}]")
# ─── Evaluation Logic ────────────────────────────────────────────────────────
def normalize_unicode(text: str) -> str:
"""Normalize Unicode subscripts/superscripts to ASCII equivalents."""
import unicodedata
# Common subscript/superscript replacements
replacements = {
'₂': '2', '₃': '3', '₄': '4', '₅': '5', '₆': '6',
'₀': '0', '₁': '1', '₇': '7', '₈': '8', '₉': '9',
'²': '2', '³': '3', '⁴': '4', '⁵': '5', '⁶': '6',
'⁰': '0', '¹': '1', '⁷': '7', '⁸': '8', '⁹': '9',
}
for old, new in replacements.items():
text = text.replace(old, new)
return text
def check_keywords(response: str, keywords: list, min_matches: int = 2) -> bool:
"""Check if response contains enough of the expected keywords.
Requires at least `min_matches` keywords to match to avoid false positives
from base models that hallucinate topic-relevant but factually wrong responses.
For short keyword lists (<=2), requires all to match.
"""
if not keywords:
return False
response_lower = normalize_unicode(response.lower())
matches = sum(1 for kw in keywords if kw in response_lower)
required = min(min_matches, len(keywords)) # Don't require more than we have
return matches >= required
def pretest_facts(facts: list) -> tuple:
"""Pre-test all facts against the model. Return (unknown, known) split."""
unknown = []
known = []
print(f"\n Pre-testing {len(facts)} facts against model...")
for i, fact in enumerate(facts):
response = daemon_query(fact.question, max_tokens=MAX_TOKENS_PRETEST)
fact.pretest_response = response
fact.pretest_known = check_keywords(response, fact.keywords)
status = "KNOWN" if fact.pretest_known else "unknown"
if (i + 1) % 10 == 0 or fact.pretest_known:
print(f" [{i+1}/{len(facts)}] {status}: {fact.question[:60]}...")
if fact.pretest_known:
known.append(fact)
else:
unknown.append(fact)
print(f" Pre-test complete: {len(unknown)} unknown, {len(known)} already known")
return unknown, known
def evaluate_recall(facts: list) -> list:
"""Post-training: test recall of each fact. Returns list of (fact, correct, response)."""
results = []
for i, fact in enumerate(facts):
response = daemon_query(fact.question, max_tokens=MAX_TOKENS_POSTTEST)
correct = check_keywords(response, fact.keywords)
results.append((fact, correct, response))
if (i + 1) % 10 == 0:
print(f" [{i+1}/{len(facts)}] recall testing...")
return results
def evaluate_general_knowledge() -> list:
"""Test general knowledge preservation."""
results = []
for item in GENERAL_KNOWLEDGE:
response = daemon_query(item["question"], max_tokens=100)
correct = check_keywords(response, item["keywords"])
results.append((item, correct, response))
return results
# ─── Statistics ──────────────────────────────────────────────────────────────
def clopper_pearson(k: int, n: int, alpha: float = 0.05) -> tuple:
"""Clopper-Pearson exact binomial confidence interval."""
if n == 0:
return (0.0, 0.0)
from scipy import stats as scipy_stats
lo = scipy_stats.beta.ppf(alpha / 2, k, n - k + 1) if k > 0 else 0.0
hi = scipy_stats.beta.ppf(1 - alpha / 2, k + 1, n - k) if k < n else 1.0
return (lo, hi)
def wilson_interval(k: int, n: int, z: float = 1.96) -> tuple:
"""Wilson score confidence interval (no scipy needed)."""
if n == 0:
return (0.0, 0.0)
p_hat = k / n
denom = 1 + z**2 / n
center = (p_hat + z**2 / (2 * n)) / denom
margin = z * math.sqrt((p_hat * (1 - p_hat) + z**2 / (4 * n)) / n) / denom
return (max(0.0, center - margin), min(1.0, center + margin))
# ─── Single Trial ────────────────────────────────────────────────────────────
def run_trial(facts: list, trial_id: int, epochs: int = TRAIN_EPOCHS) -> TrialResult:
"""Run a single trial: reset → pre-test → train → evaluate."""
print(f"\n{'='*70}")
print(f" TRIAL {trial_id}")
print(f"{'='*70}")
# 1. Reset adapter for clean slate
print(" Resetting adapter and data buffers...")
daemon_reset()
time.sleep(2)
# 2. Pre-test: confirm model doesn't know these facts
unknown_facts, known_facts = pretest_facts(facts)
if len(unknown_facts) < 10:
print(f" WARNING: Only {len(unknown_facts)} unknown facts — insufficient for evaluation")
# Still proceed but flag it
# 3. Generate training pairs from unknown facts
novel_pairs = [f.to_training_pair() for f in unknown_facts]
# 4. Calculate regularization needed for ≥33% ratio
n_reg_needed = max(1, int(len(novel_pairs) * REGULARIZATION_RATIO / (1 - REGULARIZATION_RATIO)))
n_reg_used = min(n_reg_needed, len(REGULARIZATION_PAIRS))
reg_pairs = REGULARIZATION_PAIRS[:n_reg_used]
all_pairs = novel_pairs + reg_pairs
random.shuffle(all_pairs)
print(f" Training data: {len(novel_pairs)} novel + {n_reg_used} regularization = {len(all_pairs)} total")
print(f" Regularization ratio: {n_reg_used / len(all_pairs) * 100:.1f}%")
# 5. Train
print(f" Training ({epochs} epochs max, early stopping enabled)...")
t0 = time.time()
train_result = daemon_inject_and_train(all_pairs, epochs=epochs)
train_time = time.time() - t0
print(f" Training complete: {train_time:.1f}s")
print(f" {json.dumps({k: train_result.get(k) for k in ['steps', 'final_loss', 'initial_loss', 'epochs_completed', 'early_stopped']}, default=str)}")
time.sleep(2) # Let model settle
# 6. Post-test: recall of unknown facts
print(f"\n Evaluating recall ({len(unknown_facts)} facts)...")
recall_results = evaluate_recall(unknown_facts)
recall_correct = sum(1 for _, c, _ in recall_results if c)
# 7. General knowledge preservation
print(f" Evaluating general knowledge ({len(GENERAL_KNOWLEDGE)} questions)...")
gen_results = evaluate_general_knowledge()
gen_correct = sum(1 for _, c, _ in gen_results if c)
# 8. Per-category breakdown
category_scores = {}
for fact, correct, _ in recall_results:
cat = fact.category
if cat not in category_scores:
category_scores[cat] = {"correct": 0, "total": 0}
category_scores[cat]["total"] += 1
if correct:
category_scores[cat]["correct"] += 1
result = TrialResult(
trial_id=trial_id,
n_facts_tested=len(facts),
n_confirmed_unknown=len(unknown_facts),
n_training_pairs=len(all_pairs),
n_regularization_pairs=n_reg_used,
training_steps=train_result.get("steps", 0),
training_time_s=train_time,
initial_loss=train_result.get("initial_loss", 0),
final_loss=train_result.get("final_loss", 0),
recall_correct=recall_correct,
recall_total=len(unknown_facts),
general_correct=gen_correct,
general_total=len(GENERAL_KNOWLEDGE),
category_scores=category_scores,
)
# Print trial summary
print(f"\n Trial {trial_id} Results:")
print(f" Recall: {recall_correct}/{len(unknown_facts)} ({recall_correct/max(1,len(unknown_facts))*100:.1f}%)")
print(f" General Knowledge: {gen_correct}/{len(GENERAL_KNOWLEDGE)} ({gen_correct/len(GENERAL_KNOWLEDGE)*100:.1f}%)")
print(f" Training: {result.training_steps} steps, {train_time:.1f}s, loss {result.initial_loss:.3f}{result.final_loss:.3f}")
# Print failures for debugging
failures = [(f, r) for f, c, r in recall_results if not c]
if failures:
print(f"\n Failed recalls ({len(failures)}):")
for fact, resp in failures[:10]:
print(f" Q: {fact.question[:70]}")
print(f" Expected keywords: {fact.keywords}")
print(f" Got: {resp[:100]}")
print()
gen_failures = [(item, r) for item, c, r in gen_results if not c]
if gen_failures:
print(f" General knowledge failures ({len(gen_failures)}):")
for item, resp in gen_failures:
print(f" Q: {item['question']}")
print(f" Expected: {item['keywords']}")
print(f" Got: {resp[:100]}")
return result
def run_trial_prefiltered(unknown_facts: list, trial_id: int, epochs: int = TRAIN_EPOCHS) -> TrialResult:
"""Run a trial with pre-filtered facts (already confirmed unknown). Skips pre-testing."""
print(f"\n{'='*70}")
print(f" TRIAL {trial_id}")
print(f"{'='*70}")
# 1. Reset adapter for clean slate
print(" Resetting adapter and data buffers...")
daemon_reset()
time.sleep(2)
# 2. Generate training pairs from unknown facts
novel_pairs = [f.to_training_pair() for f in unknown_facts]
# 3. Calculate regularization needed for ≥33% ratio
n_reg_needed = max(1, int(len(novel_pairs) * REGULARIZATION_RATIO / (1 - REGULARIZATION_RATIO)))
n_reg_used = min(n_reg_needed, len(REGULARIZATION_PAIRS))
reg_pairs = REGULARIZATION_PAIRS[:n_reg_used]
all_pairs = novel_pairs + reg_pairs
random.shuffle(all_pairs)
print(f" Training data: {len(novel_pairs)} novel + {n_reg_used} regularization = {len(all_pairs)} total")
print(f" Regularization ratio: {n_reg_used / len(all_pairs) * 100:.1f}%")
# 4. Train (auto_train stays off — we train explicitly via /train)
print(f" Training ({epochs} epochs max, early stopping enabled)...")
t0 = time.time()
train_result = daemon_inject_and_train(all_pairs, epochs=epochs)
train_time = time.time() - t0
print(f" Training complete: {train_time:.1f}s")
print(f" {json.dumps({k: train_result.get(k) for k in ['steps', 'final_loss', 'initial_loss', 'epochs_completed', 'early_stopped']}, default=str)}")
time.sleep(2) # Let model settle
# 5. Post-test: recall of unknown facts (auto_train disabled to avoid contamination)
daemon_set_auto_train(False)
print(f"\n Evaluating recall ({len(unknown_facts)} facts)...")
recall_results = evaluate_recall(unknown_facts)
recall_correct = sum(1 for _, c, _ in recall_results if c)
# 6. General knowledge preservation
print(f" Evaluating general knowledge ({len(GENERAL_KNOWLEDGE)} questions)...")
gen_results = evaluate_general_knowledge()
gen_correct = sum(1 for _, c, _ in gen_results if c)
# 7. Per-category breakdown
category_scores = {}
for fact, correct, _ in recall_results:
cat = fact.category
if cat not in category_scores:
category_scores[cat] = {"correct": 0, "total": 0}
category_scores[cat]["total"] += 1
if correct:
category_scores[cat]["correct"] += 1
result = TrialResult(
trial_id=trial_id,
n_facts_tested=len(unknown_facts),
n_confirmed_unknown=len(unknown_facts),
n_training_pairs=len(all_pairs),
n_regularization_pairs=n_reg_used,
training_steps=train_result.get("steps", 0),
training_time_s=train_time,
initial_loss=train_result.get("initial_loss", 0),
final_loss=train_result.get("final_loss", 0),
recall_correct=recall_correct,
recall_total=len(unknown_facts),
general_correct=gen_correct,
general_total=len(GENERAL_KNOWLEDGE),
category_scores=category_scores,
)
# Print trial summary
print(f"\n Trial {trial_id} Results:")
print(f" Recall: {recall_correct}/{len(unknown_facts)} ({recall_correct/max(1,len(unknown_facts))*100:.1f}%)")
print(f" General Knowledge: {gen_correct}/{len(GENERAL_KNOWLEDGE)} ({gen_correct/len(GENERAL_KNOWLEDGE)*100:.1f}%)")
print(f" Training: {result.training_steps} steps, {train_time:.1f}s, loss {result.initial_loss:.3f}{result.final_loss:.3f}")
# Print failures for debugging
failures = [(f, r) for f, c, r in recall_results if not c]
if failures:
print(f"\n Failed recalls ({len(failures)}):")
for fact, resp in failures[:10]:
print(f" Q: {fact.question[:70]}")
print(f" Expected keywords: {fact.keywords}")
print(f" Got: {resp[:100]}")
print()
gen_failures = [(item, r) for item, c, r in gen_results if not c]
if gen_failures:
print(f" General knowledge failures ({len(gen_failures)}):")
for item, resp in gen_failures:
print(f" Q: {item['question']}")
print(f" Expected: {item['keywords']}")
print(f" Got: {resp[:100]}")
return result
# ─── Multi-Trial Analysis ────────────────────────────────────────────────────
def run_evaluation(facts: list, n_trials: int = 3, epochs: int = TRAIN_EPOCHS):
"""Run multiple independent trials and report aggregate statistics."""
print(f"\n{'#'*70}")
print(f" STATISTICAL JIT LoRA EVALUATION")
print(f" Model: {daemon_status()['model_key']}")
print(f" Facts available: {len(facts)}")
print(f" Trials: {n_trials}")
print(f" Epochs: {epochs} (with early stopping)")
print(f" Regularization target: {REGULARIZATION_RATIO*100:.0f}%")
print(f"{'#'*70}")
# Disable auto_train during pre-testing to avoid contamination
daemon_set_auto_train(False)
# Pre-test once (base model is the same for all trials after reset)
print(f"\n === Pre-testing all {len(facts)} facts (one-time baseline) ===")
daemon_reset()
time.sleep(2)
unknown_facts, known_facts = pretest_facts(facts)
print(f"\n Baseline: {len(unknown_facts)} confirmed unknown, {len(known_facts)} already known")
print(f" Will train on {len(unknown_facts)} unknown facts across {n_trials} trials\n")
if len(unknown_facts) < 10:
print(" ERROR: Too few unknown facts for meaningful evaluation.")
print(" The model already knows most of the dataset.")
return None
results = []
for trial in range(1, n_trials + 1):
# Shuffle facts for each trial to avoid ordering effects
trial_unknown = unknown_facts.copy()
random.shuffle(trial_unknown)
result = run_trial_prefiltered(trial_unknown, trial, epochs)
results.append(result)
# ─── Aggregate Statistics ────────────────────────────────────────────
print(f"\n{'='*70}")
print(f" AGGREGATE RESULTS ({n_trials} trials)")
print(f"{'='*70}")
# Recall rates across trials
recall_rates = [r.recall_correct / max(1, r.recall_total) for r in results]
general_rates = [r.general_correct / max(1, r.general_total) for r in results]
training_times = [r.training_time_s for r in results]
training_steps_list = [r.training_steps for r in results]
n_unknown_list = [r.n_confirmed_unknown for r in results]
# Pooled counts for CI calculation
pooled_recall_k = sum(r.recall_correct for r in results)
pooled_recall_n = sum(r.recall_total for r in results)
pooled_gen_k = sum(r.general_correct for r in results)
pooled_gen_n = sum(r.general_total for r in results)
recall_ci = wilson_interval(pooled_recall_k, pooled_recall_n)
general_ci = wilson_interval(pooled_gen_k, pooled_gen_n)
print(f"\n Confirmed unknown facts per trial: {n_unknown_list}")
print(f" (facts the model verified it did NOT know before training)")
print(f"\n ┌─────────────────────────────────────────────────────────────────┐")
print(f" │ RECALL (post-training) │")
print(f" │ Pooled: {pooled_recall_k}/{pooled_recall_n} ({pooled_recall_k/max(1,pooled_recall_n)*100:.1f}%) │")
print(f" │ Per-trial rates: {[f'{r:.1%}' for r in recall_rates]}")
if n_trials > 1 and len(recall_rates) > 1:
print(f" │ Mean ± StdDev: {statistics.mean(recall_rates):.1%} ± {statistics.stdev(recall_rates):.1%}")
print(f" │ 95% CI (Wilson): [{recall_ci[0]:.1%}, {recall_ci[1]:.1%}]")
print(f" │ │")
print(f" │ GENERAL KNOWLEDGE (preservation) │")
print(f" │ Pooled: {pooled_gen_k}/{pooled_gen_n} ({pooled_gen_k/max(1,pooled_gen_n)*100:.1f}%) │")
print(f" │ Per-trial rates: {[f'{r:.1%}' for r in general_rates]}")
if n_trials > 1 and len(general_rates) > 1:
print(f" │ Mean ± StdDev: {statistics.mean(general_rates):.1%} ± {statistics.stdev(general_rates):.1%}")
print(f" │ 95% CI (Wilson): [{general_ci[0]:.1%}, {general_ci[1]:.1%}]")
print(f" │ │")
print(f" │ TRAINING │")
print(f" │ Mean time: {statistics.mean(training_times):.1f}s ± {statistics.stdev(training_times) if len(training_times) > 1 else 0:.1f}s")
print(f" │ Mean steps: {statistics.mean(training_steps_list):.0f}")
print(f" └─────────────────────────────────────────────────────────────────┘")
# Per-category aggregation
all_categories = set()
for r in results:
all_categories.update(r.category_scores.keys())
print(f"\n Per-Category Recall (pooled across trials):")
print(f" {'Category':<25} {'Correct':>8} {'Total':>8} {'Rate':>8} {'95% CI':>16}")
print(f" {'-'*25} {'-'*8} {'-'*8} {'-'*8} {'-'*16}")
for cat in sorted(all_categories):
cat_k = sum(r.category_scores.get(cat, {}).get("correct", 0) for r in results)
cat_n = sum(r.category_scores.get(cat, {}).get("total", 0) for r in results)
if cat_n > 0:
cat_ci = wilson_interval(cat_k, cat_n)
print(f" {cat:<25} {cat_k:>8} {cat_n:>8} {cat_k/cat_n:>8.1%} [{cat_ci[0]:.1%}, {cat_ci[1]:.1%}]")
# Save results to JSON
output = {
"model": daemon_status().get("model_key", "unknown"),
"n_trials": n_trials,
"epochs": epochs,
"regularization_ratio": REGULARIZATION_RATIO,
"aggregate": {
"recall": {
"pooled_correct": pooled_recall_k,
"pooled_total": pooled_recall_n,
"pooled_rate": pooled_recall_k / max(1, pooled_recall_n),
"per_trial_rates": recall_rates,
"mean": statistics.mean(recall_rates),
"stdev": statistics.stdev(recall_rates) if len(recall_rates) > 1 else 0,
"ci_95_lower": recall_ci[0],
"ci_95_upper": recall_ci[1],
},
"general_knowledge": {
"pooled_correct": pooled_gen_k,
"pooled_total": pooled_gen_n,
"pooled_rate": pooled_gen_k / max(1, pooled_gen_n),
"per_trial_rates": general_rates,
"mean": statistics.mean(general_rates),
"stdev": statistics.stdev(general_rates) if len(general_rates) > 1 else 0,
"ci_95_lower": general_ci[0],
"ci_95_upper": general_ci[1],
},
"training": {
"mean_time_s": statistics.mean(training_times),
"stdev_time_s": statistics.stdev(training_times) if len(training_times) > 1 else 0,
"mean_steps": statistics.mean(training_steps_list),
"per_trial_times": training_times,
},
},
"trials": [
{
"trial_id": r.trial_id,
"n_confirmed_unknown": r.n_confirmed_unknown,
"n_training_pairs": r.n_training_pairs,
"training_steps": r.training_steps,
"training_time_s": r.training_time_s,
"initial_loss": r.initial_loss,
"final_loss": r.final_loss,
"recall_correct": r.recall_correct,
"recall_total": r.recall_total,
"recall_rate": r.recall_correct / max(1, r.recall_total),
"general_correct": r.general_correct,
"general_total": r.general_total,
"general_rate": r.general_correct / max(1, r.general_total),
"category_scores": r.category_scores,
}
for r in results
],
}
results_path = os.path.join(os.path.dirname(__file__), "evaluation_results.json")
with open(results_path, "w") as f:
json.dump(output, f, indent=2)
print(f"\n Results saved to: {results_path}")
return output
# ─── Main ────────────────────────────────────────────────────────────────────
def main():
parser = argparse.ArgumentParser(description="Statistical JIT LoRA evaluation")
parser.add_argument("--facts-file", default=DEFAULT_FACTS_FILE,
help="Path to raw facts file (default: raw_facts_2026.txt)")
parser.add_argument("--trials", type=int, default=3,
help="Number of independent trials (default: 3)")
parser.add_argument("--max-facts", type=int, default=0,
help="Max facts to use (0 = all, default: 0)")
parser.add_argument("--epochs", type=int, default=TRAIN_EPOCHS,
help=f"Training epochs per trial (default: {TRAIN_EPOCHS})")
parser.add_argument("--seed", type=int, default=42,
help="Random seed for reproducibility (default: 42)")
args = parser.parse_args()
random.seed(args.seed)
# Verify daemon is running
try:
status = daemon_status()
if not status.get("active"):
print("ERROR: Daemon not active. Call /activate first.")
sys.exit(1)
print(f"Daemon OK: {status['model_key']}, {status.get('trainable_params', '?')} trainable params")
except Exception as e:
print(f"ERROR: Cannot reach daemon at {DAEMON_URL}: {e}")
sys.exit(1)
# Load facts
if not os.path.exists(args.facts_file):
print(f"ERROR: Facts file not found: {args.facts_file}")
print("Generate it first by running the web scraper or provide a path.")
sys.exit(1)
facts = load_facts_from_file(args.facts_file)
print(f"Loaded {len(facts)} facts from {args.facts_file}")
# Deduplicate by question
seen = set()
unique_facts = []
for f in facts:
key = f.question.lower().strip()
if key not in seen:
seen.add(key)
unique_facts.append(f)
facts = unique_facts
print(f"After dedup: {len(facts)} unique facts")
# Category distribution
cats = {}
for f in facts:
cats[f.category] = cats.get(f.category, 0) + 1
print(f"Categories: {dict(sorted(cats.items()))}")
if args.max_facts > 0 and args.max_facts < len(facts):
# Sample proportionally from each category
facts = random.sample(facts, args.max_facts)
print(f"Sampled down to {len(facts)} facts")
# Run evaluation
output = run_evaluation(facts, n_trials=args.trials, epochs=args.epochs)
# Final verdict
recall_rate = output["aggregate"]["recall"]["mean"]
gen_rate = output["aggregate"]["general_knowledge"]["mean"]
print(f"\n{'='*70}")
if recall_rate >= 0.50 and gen_rate >= 0.80:
print(f" ✓ EVALUATION PASSED")
print(f" Recall: {recall_rate:.1%} (≥50% threshold)")
print(f" General Knowledge: {gen_rate:.1%} (≥80% threshold)")
else:
print(f" ✗ EVALUATION BELOW THRESHOLD")
print(f" Recall: {recall_rate:.1%} {'✓' if recall_rate >= 0.50 else '✗ (<50%)'}")
print(f" General Knowledge: {gen_rate:.1%} {'✓' if gen_rate >= 0.80 else '✗ (<80%)'}")
print(f"{'='*70}")
if __name__ == "__main__":
main()