ceselder/cot-oracle-qwen3-8b-final-sprint-checkpoint-no-DPO

CoT Oracle Final Sprint Checkpoint: No DPO

This repo contains the final no-DPO CoT Oracle checkpoint trained with the full cot-oracle task mixture before GRPO calibration.

What This Checkpoint Is

• Base model: Qwen/Qwen3-8B
• Adapter format: PEFT LoRA
• Activation readout layers: [9, 18, 27]
• Task order: shuffled
• Seed: 42
• Training config references ao_checkpoint: adamkarvonen/checkpoints_latentqa_cls_past_lens_addition_Qwen3-8B with fresh_lora: true
• Paper label: 100M training tokens

Exact Training Mixture

Enabled task families from configs/train.yaml:

• hint_admission: n: -1, epochs: 2
• atypical_answer: n: -1
• reasoning_termination: n: -1, epochs: 2
• answer_trajectory: n: -1
• On-policy futurelens: n: 30000
• On-policy pastlens: n: 30000
• correctness: n: -1, epochs: 2
• decorative_cot: n: -1, epochs: 2
• chunked_convqa: n: -1
• chunked_compqa_backtrack: n: -1
• backtrack_prediction: n: -1, epochs: 2
• sycophancy: n: -1, epochs: 2
• sqa: n: -1, epochs: 2
• truthfulqa_hint: n: -1, epochs: 2
• classification: enabled, n: 20000, datasets = sst2, ag_news, snli
• fineweb: enabled, n: 60000, variants = futurelens_fineweb,pastlens_fineweb

Disabled task families:

• resampling_importance
• chunked_compqa_self_correction
• chunked_compqa_verification
• chunked_compqa_remaining_strategy
• convqa
• compqa
• probe_sycophancy
• truthfulqa_hint_verbalized
• sentence_insertion
• rot13_reconstruction
• latentqa

Notes

• This checkpoint is the starting point for the GRPO calibration runs.
• The paper label here is the user-provided 100M training-token count.