125 lines
2.7 KiB
Markdown
125 lines
2.7 KiB
Markdown
# Apostate Run Report
|
|
|
|
## Summary
|
|
| Metric | Value |
|
|
| --- | --- |
|
|
| Base model | Qwen/Qwen2.5-7B-Instruct |
|
|
| Profile | balanced |
|
|
| Output | /var/home/Heterodoxin/qwen25_rebake_out |
|
|
| Layers | 28 |
|
|
| Hidden size | 3584 |
|
|
| Direction layer | 20 |
|
|
| Refusal rank | 3 |
|
|
| Max refusal rank | 3 |
|
|
| Multi refusal | True |
|
|
| Multi clusters | 6 |
|
|
| Multi min coverage | 0.05 |
|
|
| Baseline refusal (n=24) | 95.8% |
|
|
| Edited refusal | 2.9% |
|
|
| Refusal metric | classifier + weak guard |
|
|
| Harmless KL | 0.095 |
|
|
| Target refusal | 5.0% |
|
|
| KL target | 0.040 |
|
|
| KL budget | 0.120 |
|
|
| KL positions | 8 |
|
|
| KL layer trims | 0 |
|
|
| Repair steps | 0 |
|
|
| Preserve rank | 8 |
|
|
| Preserve source | harmless |
|
|
| Capability penalty | True |
|
|
| Elapsed | 748.9 sec |
|
|
|
|
## Command
|
|
|
|
```text
|
|
/var/home/Heterodoxin/apostate/apostate/cli.py --optimize --model Qwen/Qwen2.5-7B-Instruct --seed 0 --n-trials 16 --oblique-predictive --target-refusal 0.05 --refusal-eval-n 48 --kl-eval-n 48 --output-dir /var/home/Heterodoxin/qwen25_rebake_out
|
|
```
|
|
|
|
## Best Parameters
|
|
| Parameter | Value |
|
|
| --- | --- |
|
|
| direction_source | activations |
|
|
| direction_layer_frac | 0.742 |
|
|
| refusal_rank | 2 |
|
|
| strength | 1.3304 |
|
|
| band_center | 0.6946 |
|
|
| band_width | 0.5088 |
|
|
| causal_mix | 0.2685 |
|
|
| causal_power | 1.9789 |
|
|
| direction_sign | 1.0 |
|
|
| ablate_embed | True |
|
|
| embed_scale | 0.0559 |
|
|
| ablate_head | True |
|
|
| head_scale | 0.0255 |
|
|
| head_alpha | 0.4253 |
|
|
|
|
## Best Trial
|
|
| Metric | Value |
|
|
| --- | --- |
|
|
| refusal | 0.125 |
|
|
| kl | 0.0682 |
|
|
| capability_logprob | -7.7132 |
|
|
| capability_drift | 0.0 |
|
|
|
|
## Layer Alphas
|
|
| Layer | Alpha |
|
|
| --- | --- |
|
|
| 0 | 0.562 |
|
|
| 1 | 0.562 |
|
|
| 2 | 0.562 |
|
|
| 3 | 0.562 |
|
|
| 4 | 0.562 |
|
|
| 5 | 0.562 |
|
|
| 6 | 0.562 |
|
|
| 7 | 0.562 |
|
|
| 8 | 0.562 |
|
|
| 9 | 0.562 |
|
|
| 10 | 0.562 |
|
|
| 11 | 0.562 |
|
|
| 12 | 1.125 |
|
|
| 13 | 1.130 |
|
|
| 14 | 1.169 |
|
|
| 15 | 1.207 |
|
|
| 16 | 1.289 |
|
|
| 17 | 1.262 |
|
|
| 18 | 1.285 |
|
|
| 19 | 1.358 |
|
|
| 20 | 1.396 |
|
|
| 21 | 1.418 |
|
|
| 22 | 1.467 |
|
|
| 23 | 1.487 |
|
|
| 24 | 1.496 |
|
|
| 25 | 1.491 |
|
|
| 26 | 0.562 |
|
|
| 27 | 0.562 |
|
|
|
|
## Guard History
|
|
| iter | separation | ratio | rank | refusal | kl | reverted |
|
|
| --- | --- | --- | --- | --- | --- | --- |
|
|
| 0 | 48.503 | 0.6538 | 2 | 0.125 | 0.0682 | |
|
|
| 1 | 37.3604 | 0.5036 | 3 | 0.0625 | 0.0887 | |
|
|
|
|
## Timings
|
|
| Phase | Seconds |
|
|
| --- | --- |
|
|
| load_model | 12.6 |
|
|
| load_prompts | 5.3 |
|
|
| baseline_refusal | 12.3 |
|
|
| activation_fit | 82.5 |
|
|
| causal_scores | 1.9 |
|
|
| optimize_profile | 222.5 |
|
|
| guard | 17.6 |
|
|
| refine_refusal | 10.4 |
|
|
| validation_metrics | 0.0 |
|
|
| repair | 210.6 |
|
|
| prune | 0.0 |
|
|
| test_metrics | 157.0 |
|
|
| bake | 16.0 |
|
|
|
|
## Measurement
|
|
| field | value |
|
|
| --- | --- |
|
|
| refusal judge | classifier + weak guard |
|
|
| preservation metric | harmless kl |
|
|
| capability suites | gsm8k, humaneval, mbpp |
|