Upload add_sub_baseline_100K

add_sub_baseline_100K/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b37212c1e47926f053e7948899c776f0b4b971636d7c2b42f34683c062cf1fd1
+oid sha256:0169da874edcb1493b3da2744c9d0e42387fe999477b4cfec1e956d819d1538d
 size 650266922

add_sub_baseline_100K/train_config.json

@@ -1,35 +1,84 @@
 {
-  "mode": "baseline",
-  "ops": "add_sub",
-  "n_digits": 6,
-  "n_layer": 2,
-  "n_head": 3,
-  "n_embd": 510,
-  "abs_vocab": 0,
+  "num_rollouts": 4,
   "K": 4,
+  "max_iterations": 2,
+  "memory_span_abs": 1792,
+  "memory_span_traj": 1792,
+  "temperature": 1.0,
+  "ar_search": false,
+  "response_only_abs": false,
   "alpha_info_gain": 10.0,
   "alpha_abs": 0.1,
   "alpha_soft_zipf": 1.0,
+  "alpha_ortho": 0.0,
+  "alpha_anchor": 0.0,
+  "alpha_jacobi": 0.0,
+  "decay": 0.8,
+  "target_vocab_util": 0.8,
+  "min_abs_ppl": 0.0,
+  "zipf_alpha": 1.0,
+  "lr": 4e-05,
+  "emb_lr_mult": 1.0,
+  "weight_decay": 0.01,
+  "warmup_steps": 937,
+  "cooldown_frac": 0.4,
+  "max_grad_norm": 1.0,
+  "vq_abs_pretrain_steps": 0,
+  "vq_abs_pretrain_lr": 0.001,
+  "vq_abs_pretrain_layer": -1,
+  "vq_abs_pretrain_batch_size": 256,
+  "vq_abs_pretrain_target_vectors": 20000,
   "batch_size": 64,
+  "gradient_accumulation_steps": 1,
   "num_epochs": 20,
-  "dataset_size": 100000,
-  "lr": 8e-05,
+  "emb_warmup_steps": 0,
+  "log_every": 50,
+  "eval_every": 1562,
+  "save_every": 999999,
+  "eval_samples": 100,
   "output_dir": "ckpt/sweep/add_sub_baseline_100K",
+  "eval_K": 4,
+  "alpha_traj": 0.0,
+  "corrupt_method": "shuffle",
+  "corrupt_ratio": 0.3,
+  "alpha_contrastive": 1.0,
+  "gamma_contrastive": 0.5,
+  "alpha_masked_traj": 0.0,
+  "mask_nl_ratio": 0.3,
+  "mask_nl_mode": "fixed",
+  "mask_nl_fixed_id": 0,
+  "use_ste": true,
+  "n_inner": 1,
+  "random_K": null,
+  "strip_suffix": null,
+  "compress_prefix": null,
+  "random_mem_span": null,
+  "warmup_ratio": 0.03,
+  "beta2": 0.999,
+  "seed": 42,
+  "n_digits": 6,
+  "n_layer": 2,
+  "n_head": 3,
+  "n_embd": 510,
+  "ops": "add_sub",
+  "abs_vocab": 0,
+  "dataset_size": 100000,
+  "mode": "baseline",
   "device": "cuda",
   "push_to_hub": true,
   "no_wandb": false,
   "n_params": 162490082,
   "run_name": "add_sub_baseline_100K",
-  "git_commit": "800625019270114adcda289bbd550c4f1109a514",
-  "timestamp": "2026-04-12T01:58:11.825040+00:00",
+  "git_commit": "f447da529caceac8c7d256cbb2cd185cbc50feac",
+  "timestamp": "2026-04-12T09:39:33.458507+00:00",
   "tokenizer": "Qwen/Qwen3-0.6B",
   "dataset_repo": "thoughtworks/arithmetic-sorl-data",
   "dataset_config": "add_sub_6digit",
   "model_repo": "thoughtworks/arithmetic-sorl",
   "trainer_version": "sft",
-  "wandb_run_id": "ngy2a12o",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/ngy2a12o",
-  "final_accuracy": 0.9964285714285714,
-  "sft_accuracy": 0.9964285714285714,
+  "wandb_run_id": "f485r2vd",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/f485r2vd",
+  "final_accuracy": 1.0,
+  "sft_accuracy": 1.0,
   "eval_method": "ArithmeticEvaluator"
 }