| # CTX next-scene: NK2E stock recipe on the 1255-pair curated contexto dataset | |
| # (original captions = pure description of image B, same data condition as the | |
| # validated omini_grounded probe). Template replaces the edit-instruction wrapper: | |
| # these captions are next-scene descriptions, not minimal-edit instructions. | |
| method: incontext | |
| checkpoint_env: OSS_RAW | |
| precision: nf4 # 32GB GPU; encoder stays resident | |
| grad_checkpointing: true | |
| lora_rank: 64 | |
| lora_alpha: 64 | |
| lora_dropout: 0.0 | |
| lora_targets: ["attn.wq", "attn.wk", "attn.wv", "attn.wo", "mlp.gate", "mlp.up", "mlp.down"] | |
| lora_include_prefix: "blocks." | |
| cache_dir: /workspace/datasets/ctx_original/cache_nk2e_512 | |
| size: 512 | |
| t_logitnorm_mean: 0.0 | |
| t_logitnorm_std: 1.0 | |
| prompt_dropout: 0.1 | |
| instruction_template: "Create the next scene. {instruction}" | |
| batch_size: 1 | |
| grad_accum: 1 | |
| lr: 1.0e-4 | |
| weight_decay: 0.0 | |
| max_steps: 2000 # 2000 updates reais (batch 1, sem accum — pedido do usuário) | |
| warmup_steps: 0 | |
| use_adam8bit: true | |
| seed: 0 | |
| out_dir: /workspace/checkpoints/nk2e_ctx | |
| save_every: 250 | |
| log_every: 100 | |