ceselder/cot-oracle-paper-ablation-ours-1layer

CoT Oracle Paper Ablation: Ours, 1 Layer

This repo contains the 1-layer paper ablation for the CoT Oracle recipe: on-policy lens tasks, chunked ConvQA, FineWeb lens readouts, and classification, without LatentQA.

What This Checkpoint Is

• Base model: Qwen/Qwen3-8B
• Adapter format: PEFT LoRA
• Activation readout layers: [18]
• Task order: shuffled
• Seed: 42
• Planned budget: 50M input tokens
• Paper label: 22.5M logged training tokens

Exact Training Mixture

• On-policy futurelens: enabled, n: 30000
• On-policy pastlens: enabled, n: 30000
• chunked_convqa: enabled, n: -1 (all available examples)
• classification: enabled, n: 20000, datasets = sst2, ag_news, snli
• fineweb: enabled, n: 60000, variants = futurelens_fineweb,pastlens_fineweb
• latentqa: disabled
• All other tasks in configs/train.yaml: disabled

Notes

• This is the 1-layer counterpart to the 3-layer paper ablations.
• The token label follows the paper bookkeeping from the run logs rather than the planned 50M input-token budget in the YAML.