Add training_config.json

training_config.json

@@ -0,0 +1,26 @@
+{
+  "model_name": "Qwen/Qwen3.5-2B",
+  "hf_repo_id": "farpluto/doc-to-lora-niah-qwen3.5-2B",
+  "lora_r": 8,
+  "lora_alpha": 8.0,
+  "target_module": "down_proj",
+  "latent_dim": 512,
+  "n_percv_blocks": 8,
+  "lr": 8e-05,
+  "weight_decay": 0.01,
+  "grad_accum": 2,
+  "max_steps": 1500,
+  "warmup_steps": 100,
+  "max_grad_norm": 1.0,
+  "l1_coef": 0.05,
+  "use_kl_loss": false,
+  "kl_coef": 1.0,
+  "eval_every": 200,
+  "save_every": 1000,
+  "ctx_min_len": 32,
+  "ctx_max_len": 256,
+  "n_train": 80000,
+  "n_eval": 300,
+  "seed": 42,
+  "out_dir": "d2l_ckpts"
+}