medium fw qa 3 + fallback to manual cache dataset

configs/pretrain_all_xl_3_medium.yaml

@@ -0,0 +1,60 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+label_names: ["labels"]
+# eval_on_start: True
+# eval_strategy: "steps"
+# eval_steps: 500
+# save_strategy: "no"
+# # save_steps: 500
+# logging_strategy: "steps"
+# logging_steps: 100
+# use_liger_kernel: true
+# remove_unused_columns: false
+
+# needed to avoid OOM by compute the metrics batch by batch
+# w/o this the trainer stores logits of all sample in memory...
+# batch_eval_metrics: true
+
+per_device_train_batch_size: 8
+per_device_eval_batch_size: 8
+max_val_samples_per_ds: 1000
+# optim: schedule_free_adamw
+
+learning_rate: 0.00004
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 5
+weight_decay: 0.01
+#
+warmup_steps: 100
+
+dataloader_prefetch_factor: 8
+dataloader_num_workers: 8
+# LoRA
+lora_r: 8
+lora_dropout: 0.0
+target_modules:
+  - down_proj
+# data
+train_ds_names:
+  - fw_qa_3_medium # ~ 130M?
+  - ctx_qa # 300k
+  - pwc # 240k
+  - hotpot_qa # 90k
+  - squad # 90k
+  - drop # 77k
+  - narrativeqa # 40k
+  - quoref # 11k
+  - ropes # 11k
+  - synthetic_convqa # 40k
+
+val_ds_names:
+  - fw_qa_3
+  - fw_qa_xl
+  - ctx_qa
+  - pwc
+  - hotpot_qa
+  - squad
+
+load_best_model_at_end: true
+metric_for_best_model: eval_pwc_loss

configs/pretrain_all_xl_3_mini.yaml

@@ -37,7 +37,7 @@ target_modules:
   - down_proj
 # data
 train_ds_names:
-  - fw_qa_3_mini # ~ 267M
+  - fw_qa_3_mini # 100k
   - ctx_qa # 300k
   - pwc # 240k
   - hotpot_qa # 90k