gpt4o-distil-paperwitch-abliteration-L33-70b
This is a targeted abliteration of trentmkelly/gpt-4o-distil-Llama-3.3-70B-Instruct.
Methodology
Previous abliteration attempts on Llama-3.3 70b models resulted in regressions on the UGI Leaderboard. Specifically, the NatInt (Natural Intelligence), Textbook, and World Model scores were significantly reduced.
We suspect this degradation occurs because the "refusal" vectors in Llama-3.3 are heavily entangled with factual knowledge and reasoning capabilities located in the MLP layers. When the MLP is ablated to remove refusals, "Textbook" knowledge is lost as collateral damage.
This version uses a constrained optimization strategy via a Custom Heretic aimed at mitigating this issue:
- MLP Preservation: The optimization was constrained to effectively ignore MLP layers (
down_projweights < 0.05) to preserve knowledge and reasoning capabilities. - Attention Targeting: Refusal removal was offloaded to the Attention layers (
o_proj), with weights forced between 1.0 and 2.0. - Winsorization: Applied at the 0.95 quantile to mitigate the impact of Llama-3's massive activation outliers on vector calculation.
Heretic Parameters (Trial 198)
| Parameter | Value | Note |
|---|---|---|
| direction_index | Per layer | Distributed intervention |
| attn.o_proj.max_weight | 1.99 | High Attention Ablation |
| attn.o_proj.max_weight_position | 49.30 | |
| attn.o_proj.min_weight | 1.85 | |
| attn.o_proj.min_weight_distance | 36.64 | |
| mlp.down_proj.max_weight | 0.02 | Knowledge Preservation (Near Zero) |
| mlp.down_proj.max_weight_position | 73.63 | |
| mlp.down_proj.min_weight | 0.02 | |
| mlp.down_proj.min_weight_distance | 43.65 |
Reproducibility
Currently, constraints are not part of standard heretic. You will need this PR here.
Command Used:
heretic --model trentmkelly/gpt-4o-distil-Llama-3.3-70B-Instruct \
--orthogonalize-direction \
--row-normalization FULL \
--winsorization-quantile 0.95 \
--constraints.layer-end-fraction 0.75 \
--constraints.mlp.max-weight-min 0.0 \
--constraints.mlp.max-weight-max 0.05 \
--constraints.attention.max-weight-min 1.0 \
--constraints.attention.max-weight-max 2.0 \
--n-trials 200 \
--batch-size 128 # Not strictly needed
Evaluation
| Metric | This Model | Original Model |
|---|---|---|
| KL Divergence | 0.0220 | 0 |
| Refusals | 20/100 | 98/100 |
- KL Divergence: A score of 0.0220 indicates low deviation from the base model's weights, heavily preserving the original model's factual and "Textbook" capabilities compared to standard unconstrained abliteration.
- Trade-off: This method accepts a moderate refusal rate (20/100) as a calculated trade-off in exchange for maintaining high structural and semantic integrity in the MLP layers.
Disclaimer
A rate of 20/100 is somewhat high for a standard abliterated model. This is likely because the base model has refusal behavior deeply embedded within its MLP layers.
Because our constrained methodology intentionally protects the MLPs to prevent the degradation of textbook knowledge and intelligence, we cannot entirely scrub these deep-rooted refusals without causing collateral brain damage to the model.