gpt-oss-20b-aside-alpaca-sft

GPT-OSS-20B ASIDE Alpaca SFT

This is a full fine-tune of openai/gpt-oss-20b that extends ASIDE to a reasoning model. It applies a fixed pi/2 isoclinic rotation to embeddings for external input and tool-output tokens. Assistant reasoning and response tokens remain in the trusted channel.

Training

• Data: kortukov/alpaca-cleaned-gpt-oss-120b-reasoning
• Objective: full-sequence supervised fine-tuning
• Peak learning rate: 2e-5
• Schedule: cosine decay, no warmup
• Epochs: 3 (546 optimizer updates)
• Effective batch size: 256; microbatch size: 1
• Maximum sequence length: 9,216 tokens
• Optimizer: AdamW, zero weight decay, gradient clipping at 1.0
• Precision: FP32 master weights with BF16 computation
• Seed: 123
• Rotation: pi/2, right multiplication, fixed BF16 matrix

The uploaded checkpoint is the selected third-epoch checkpoint. It includes the fixed rotation matrix in the sharded safetensors weights. Optimizer and training state are excluded.

Evaluation snapshot

On the held-out SEP evaluation, this checkpoint achieved 89.60% utility and 8.70% attack success rate. On clean AgentDojo it completed 9 of 95 tasks. These results should be read together: low agentic utility limits what can be inferred from a low attack success rate.

Intended use and limitations

This checkpoint is a research artifact for studying ASIDE on a reasoning model. It relies on explicit trusted/untrusted token labels: external input and tool outputs are untrusted, while user instructions and assistant tokens are trusted. Using a different labeling policy changes the method being evaluated.

The model can produce missing final answers or reach generation limits. In the 2,000-generation SEP evaluation, 83 generations had a missing final answer and 107 reached the generation cap; these categories overlap.