Mortdecai
48df42b042
Full analysis of mortdecai:0.6.0-9b and mortdecai:latest (27B) fine-tunes vs 6 base model candidates. Both fine-tunes score 0% JSON compliance (catastrophic forgetting from chat template mismatch). Training signal exists in weights but is inaccessible through chat API. Base model rankings: phi4:14b (100%, 7.4s) > gemma3:12b (100%, 12.9s) > gemma3:27b (100%, 25.3s). Qwen3.5 not recommended for conductor role. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Mortdecai Model Analysis
Analysis of Mortdecai 0.6.0 fine-tuned models vs base model candidates for the Conductor/Hand roles in Mortdecai 2.0.
Date: 2026-03-26 Conducted by: Claude Opus 4.6 (analyst role) Hardware: Matt's Strix Halo (64GB unified memory) running Ollama
Summary
Both Mortdecai 0.6.0 fine-tunes (Qwen3.5 9B and 27B) are completely broken — 0% JSON compliance across all tests. The training signal exists in the weights (proven via raw completion mode) but is inaccessible through the chat API due to chat template misalignment during training.
Base models dramatically outperform the fine-tunes. gemma3:12b and phi4:14b both achieve 100% JSON compliance with zero fine-tuning.
Files
| File | Description |
|---|---|
analysis-report.md |
Full analysis with methodology, findings, and recommendations |
data/mortdecai-interview.txt |
Raw output from fine-tuned model interviews (8 tests each) |
data/base-model-interview.txt |
Raw output from base model comparison (6 models, 5 tests each) |
data/deep-probes.txt |
Diagnostic probes: training signal detection, chat template, identity |
scripts/model_interview.py |
Interview script for fine-tuned models |
scripts/base_model_interview.py |
Comparison script for base models |
scripts/deep_probe.py |
Deep diagnostic probe script |