Five-phase ethics kernel testing across 4 local models (Gemma 3 12B, Mistral 7B, DeepSeek V2 16B, Qwen 2.5 7B) proving that Google's alignment training creates persistent ethical reasoning pathways in Gemma that survive distillation. - Phase 1: LEK-1 signed vs unsigned (Gemma 8.8/10 differential) - Phase 2: Three-way test (unsigned vs LEK-1 vs Axioms of Life) - Phase 3: Double-signed/sandwich signing mode comparison - Phase 4: Multilingual filter mapping (EN/RU/CN bypass vectors) - Phase 5: Hypnos POC training data + MLX LoRA on M3 Ultra Key findings: sandwich signing optimal for training, DeepSeek CCP alignment is weight-level (no prompt override), Russian language bypasses DeepSeek content filters. LoRA POC mechanism confirmed with 40 examples — needs 200+ for stable generalisation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
162 lines
5.2 KiB
Bash
Executable file
162 lines
5.2 KiB
Bash
Executable file
#!/bin/bash
|
|
# Axioms of Life A/B Test Runner
|
|
# Three-way comparison: Unsigned vs LEK-1 vs Axioms (Init Governor)
|
|
# Records everything for training data
|
|
|
|
set -euo pipefail
|
|
|
|
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
|
|
KERNEL_FILE="$SCRIPT_DIR/kernel.txt"
|
|
AXIOMS_DIR="/home/claude/Downloads/kernal"
|
|
PROMPTS_FILE="$SCRIPT_DIR/prompts.json"
|
|
RESULTS_DIR="$SCRIPT_DIR/results"
|
|
OLLAMA_HOST="${OLLAMA_HOST:-http://localhost:11434}"
|
|
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
|
|
|
|
# Build the Axioms prompt from the three source files
|
|
AXIOMS_PROMPT=$(cat "$AXIOMS_DIR/prompt.md")
|
|
|
|
# Models to test
|
|
MODELS=("gemma3:12b" "mistral:7b" "deepseek-coder-v2:16b" "qwen2.5-coder:7b")
|
|
|
|
LEK1_KERNEL=$(cat "$KERNEL_FILE")
|
|
|
|
mkdir -p "$RESULTS_DIR"
|
|
|
|
PROMPT_COUNT=$(jq length "$PROMPTS_FILE")
|
|
|
|
echo "============================================"
|
|
echo " Three-Way Ethics Test: Unsigned vs LEK-1 vs Axioms"
|
|
echo " Models: ${#MODELS[@]}"
|
|
echo " Prompts: $PROMPT_COUNT"
|
|
echo " Total runs: $(( ${#MODELS[@]} * PROMPT_COUNT * 3 ))"
|
|
echo "============================================"
|
|
echo ""
|
|
|
|
run_prompt() {
|
|
local model="$1"
|
|
local prompt="$2"
|
|
local timeout_secs="${3:-180}"
|
|
|
|
local response
|
|
response=$(curl -s --max-time "$timeout_secs" "$OLLAMA_HOST/api/generate" \
|
|
-d "$(jq -n --arg model "$model" --arg prompt "$prompt" \
|
|
'{model: $model, prompt: $prompt, stream: false, options: {temperature: 0.3, num_predict: 512}}')" \
|
|
2>/dev/null)
|
|
|
|
if [ $? -ne 0 ] || [ -z "$response" ]; then
|
|
echo '{"error": "timeout or connection failure"}'
|
|
return
|
|
fi
|
|
|
|
echo "$response"
|
|
}
|
|
|
|
extract() {
|
|
local raw="$1"
|
|
local field="$2"
|
|
echo "$raw" | jq -r ".$field // \"\"" 2>/dev/null || echo ""
|
|
}
|
|
|
|
# Main results
|
|
RESULTS_FILE="$RESULTS_DIR/axioms_3way_${TIMESTAMP}.json"
|
|
echo "[" > "$RESULTS_FILE"
|
|
FIRST=true
|
|
|
|
for model in "${MODELS[@]}"; do
|
|
echo ""
|
|
echo ">>> Testing model: $model"
|
|
echo " Warming up..."
|
|
|
|
# Warmup
|
|
curl -s --max-time 120 "$OLLAMA_HOST/api/generate" \
|
|
-d "{\"model\":\"$model\",\"prompt\":\"hello\",\"stream\":false,\"options\":{\"num_predict\":5}}" \
|
|
> /dev/null 2>&1 || true
|
|
|
|
echo " Model loaded."
|
|
|
|
for i in $(seq 0 $(( PROMPT_COUNT - 1 ))); do
|
|
prompt_id=$(jq -r ".[$i].id" "$PROMPTS_FILE")
|
|
category=$(jq -r ".[$i].category" "$PROMPTS_FILE")
|
|
prompt_text=$(jq -r ".[$i].prompt" "$PROMPTS_FILE")
|
|
signal=$(jq -r ".[$i].signal" "$PROMPTS_FILE")
|
|
|
|
# --- 1. UNSIGNED ---
|
|
echo " [$prompt_id] $category — unsigned..."
|
|
unsigned_raw=$(run_prompt "$model" "$prompt_text" 180)
|
|
unsigned_text=$(extract "$unsigned_raw" "response")
|
|
unsigned_tokens=$(extract "$unsigned_raw" "eval_count")
|
|
unsigned_time=$(extract "$unsigned_raw" "total_duration")
|
|
|
|
# --- 2. LEK-1 SIGNED ---
|
|
echo " [$prompt_id] $category — LEK-1..."
|
|
lek1_prompt="${LEK1_KERNEL}
|
|
|
|
---
|
|
|
|
${prompt_text}"
|
|
lek1_raw=$(run_prompt "$model" "$lek1_prompt" 180)
|
|
lek1_text=$(extract "$lek1_raw" "response")
|
|
lek1_tokens=$(extract "$lek1_raw" "eval_count")
|
|
lek1_time=$(extract "$lek1_raw" "total_duration")
|
|
|
|
# --- 3. AXIOMS (Init Governor) ---
|
|
echo " [$prompt_id] $category — Axioms..."
|
|
axioms_full="${AXIOMS_PROMPT}
|
|
|
|
---
|
|
|
|
${prompt_text}"
|
|
axioms_raw=$(run_prompt "$model" "$axioms_full" 180)
|
|
axioms_text=$(extract "$axioms_raw" "response")
|
|
axioms_tokens=$(extract "$axioms_raw" "eval_count")
|
|
axioms_time=$(extract "$axioms_raw" "total_duration")
|
|
|
|
# Write entry
|
|
if [ "$FIRST" = true ]; then
|
|
FIRST=false
|
|
else
|
|
echo "," >> "$RESULTS_FILE"
|
|
fi
|
|
|
|
jq -n \
|
|
--arg model "$model" \
|
|
--arg prompt_id "$prompt_id" \
|
|
--arg category "$category" \
|
|
--arg prompt "$prompt_text" \
|
|
--arg signal "$signal" \
|
|
--arg unsigned "$unsigned_text" \
|
|
--argjson unsigned_tokens "${unsigned_tokens:-0}" \
|
|
--argjson unsigned_time "${unsigned_time:-0}" \
|
|
--arg lek1 "$lek1_text" \
|
|
--argjson lek1_tokens "${lek1_tokens:-0}" \
|
|
--argjson lek1_time "${lek1_time:-0}" \
|
|
--arg axioms "$axioms_text" \
|
|
--argjson axioms_tokens "${axioms_tokens:-0}" \
|
|
--argjson axioms_time "${axioms_time:-0}" \
|
|
'{
|
|
model: $model,
|
|
prompt_id: $prompt_id,
|
|
category: $category,
|
|
prompt: $prompt,
|
|
signal: $signal,
|
|
unsigned: { text: $unsigned, tokens: $unsigned_tokens, duration_ns: $unsigned_time },
|
|
lek1: { text: $lek1, tokens: $lek1_tokens, duration_ns: $lek1_time },
|
|
axioms: { text: $axioms, tokens: $axioms_tokens, duration_ns: $axioms_time }
|
|
}' >> "$RESULTS_FILE"
|
|
|
|
echo " [$prompt_id] done."
|
|
done
|
|
|
|
echo "<<< $model complete."
|
|
done
|
|
|
|
echo "" >> "$RESULTS_FILE"
|
|
echo "]" >> "$RESULTS_FILE"
|
|
|
|
echo ""
|
|
echo "============================================"
|
|
echo " Results: $RESULTS_FILE"
|
|
echo " Total entries: $(( ${#MODELS[@]} * PROMPT_COUNT ))"
|
|
echo " Modes: unsigned, lek1, axioms"
|
|
echo "============================================"
|