This is the adapter for Chimpanzee-1.1, a model custom-trained for ARC-AGI 3. To use the adapter, you will need the Qwen3.6-27b base model.
Chimpanzee-1.1 Training
Chimpanzee-1.1 was trained using multimodal DPO on unquantized Qwen3.6-27B. Training followed a cumulative RPS curriculum:
| Stage | Dataset | Pairs | Learning Rate |
|---|---|---|---|
| Stage 1 | ARC-Witness Level 1 | 75 | 3.333333333333e-6 |
| Stage 1C | ARC-Interactive Level 1 | 75 | 3.333333333333e-6 |
| Stage 2 | ARC-Witness later levels | 144 | 5e-7 |
| Stage 2C | ARC-Interactive Levels 2+ | 286 | 5e-7 |
Stage 2 and Stage 2C use 15% of the high-stage learning rate, implementing the lower-plasticity portion of the RPS curriculum. Each stage initializes from the cumulative adapter produced by the preceding stage.
Training used one epoch per stage, a constant-with-warmup scheduler, a 5% warmup ratio, DPO beta 0.1, and RPO alpha 1.0. LoRA was applied to all linear LLM modules with rank 64, alpha 128, and dropout 0.05. The vision transformer and vision-language aligner remained frozen.
Training used bfloat16 precision, SDPA, reentrant gradient checkpointing, and DeepSpeed ZeRO-3 across two 96 GB H20 GPUs. The per-device batch size was 1, gradient accumulation was 1, and the effective global batch size was 2. The maximum branch length was 11,000 tokens, and the maximum image size was 262,144 pixels.
The final Stage 2C adapter is cumulative and contains the effects of all four curriculum stages.