Upload ctx.yaml with huggingface_hub
Browse files
ctx.yaml
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# CTX next-scene: NK2E stock recipe on the 1255-pair curated contexto dataset
|
| 2 |
+
# (original captions = pure description of image B, same data condition as the
|
| 3 |
+
# validated omini_grounded probe). Template replaces the edit-instruction wrapper:
|
| 4 |
+
# these captions are next-scene descriptions, not minimal-edit instructions.
|
| 5 |
+
method: incontext
|
| 6 |
+
checkpoint_env: OSS_RAW
|
| 7 |
+
precision: nf4 # 32GB GPU; encoder stays resident
|
| 8 |
+
grad_checkpointing: true
|
| 9 |
+
|
| 10 |
+
lora_rank: 64
|
| 11 |
+
lora_alpha: 64
|
| 12 |
+
lora_dropout: 0.0
|
| 13 |
+
lora_targets: ["attn.wq", "attn.wk", "attn.wv", "attn.wo", "mlp.gate", "mlp.up", "mlp.down"]
|
| 14 |
+
lora_include_prefix: "blocks."
|
| 15 |
+
|
| 16 |
+
cache_dir: /workspace/datasets/ctx_original/cache_nk2e_512
|
| 17 |
+
size: 512
|
| 18 |
+
|
| 19 |
+
t_logitnorm_mean: 0.0
|
| 20 |
+
t_logitnorm_std: 1.0
|
| 21 |
+
prompt_dropout: 0.1
|
| 22 |
+
instruction_template: "Create the next scene. {instruction}"
|
| 23 |
+
|
| 24 |
+
batch_size: 1
|
| 25 |
+
grad_accum: 4 # effective batch 4 (same as the validated probe)
|
| 26 |
+
lr: 1.0e-4
|
| 27 |
+
weight_decay: 0.0
|
| 28 |
+
max_steps: 8000 # micro-steps; /4 = 2000 optimizer steps (pedido do usuário)
|
| 29 |
+
warmup_steps: 0
|
| 30 |
+
use_adam8bit: true
|
| 31 |
+
seed: 0
|
| 32 |
+
|
| 33 |
+
out_dir: /workspace/checkpoints/nk2e_ctx
|
| 34 |
+
save_every: 1000 # micro-steps; = every 250 optimizer steps
|
| 35 |
+
log_every: 100
|