Add lineage metadata batch 1

README.md

@@ -0,0 +1,7 @@
+# think.nano
+
+Durable model artifacts for lineage-aware nanochat experiments.
+
+Each base experiment owns its tokenizer and base checkpoints. SFT runs are
+nested under their exact base parent, and post-training runs are nested under
+their exact SFT parent. Legacy paths are retained under `archive/pre-lineage-v1`.

experiments/think-d12-r11.25/config.json

@@ -0,0 +1,49 @@
+{
+  "schema_version": 1,
+  "stage": "base",
+  "experiment_id": "think-d12-r11.25",
+  "dataset": {
+    "adapter": "parquet_shards",
+    "repo": "jbduran/think-dataset",
+    "revision": "main",
+    "validation_shard": 472,
+    "num_train_shards": 24,
+    "download_workers": 4
+  },
+  "tokenizer": {
+    "mode": "train",
+    "max_chars": 2000000000,
+    "doc_cap": 10000,
+    "vocab_size": 32768
+  },
+  "pretokenize": {
+    "enabled": true,
+    "slack": 1.03,
+    "val_tokens": 20971520,
+    "shard_tokens": 100000000,
+    "tokenizer_threads": 8
+  },
+  "training": {
+    "depth": 12,
+    "scaling_params": 110100912,
+    "target_param_data_ratio": 11.25,
+    "window_pattern": "L",
+    "device_batch_size": 16,
+    "total_batch_size": 524288,
+    "save_every": 500,
+    "eval_every": 250,
+    "eval_tokens": 2097152,
+    "core_metric_every": -1,
+    "sample_every": -1
+  },
+  "artifacts": {
+    "repo": "jbduran/think.nano"
+  },
+  "wandb": {
+    "entity": "jbduran-thinkingmachinesncsu",
+    "project": "think.nano",
+    "name": "think-d12-r11.25",
+    "group": "think-d12",
+    "tags": ["think-dataset", "d12", "ratio11.25"]
+  }
+}

experiments/think-d12-r11.25/run.json

@@ -0,0 +1,6 @@
+{
+  "experiment_id": "think-d12-r11.25",
+  "stage": "base",
+  "wandb_run_id": null,
+  "migration_note": "Migrated from the pre-lineage repository layout."
+}

experiments/think-d12-r11.25/sft/smoltalk-mmlu3-gsm8k4-v1/config.json

@@ -0,0 +1,36 @@
+{
+  "schema_version": 1,
+  "stage": "sft",
+  "experiment_id": "smoltalk-mmlu3-gsm8k4-v1",
+  "parent": {
+    "base_experiment_id": "think-d12-r11.25",
+    "checkpoint_step": 2362
+  },
+  "data": {
+    "recipe": "nanochat-default",
+    "mmlu_epochs": 3,
+    "gsm8k_epochs": 4
+  },
+  "training": {
+    "num_iterations": -1,
+    "device_batch_size": 8,
+    "eval_every": -1,
+    "chatcore_every": -1,
+    "save_every": 200
+  },
+  "artifacts": {
+    "repo": "jbduran/think.nano"
+  },
+  "wandb": {
+    "enabled": false,
+    "entity": "jbduran-thinkingmachinesncsu",
+    "project": "think.nano",
+    "name": "smoltalk-mmlu3-gsm8k4-v1",
+    "group": "think-d12-r11.25",
+    "tags": ["sft", "smoltalk", "mmlu3", "gsm8k4"]
+  },
+  "historical": {
+    "lineage_inferred_from_checkpoint_metadata": true,
+    "wandb_was_disabled": true
+  }
+}

experiments/think-d12-r11.25/sft/smoltalk-mmlu3-gsm8k4-v1/run.json

@@ -0,0 +1,6 @@
+{
+  "experiment_id": "smoltalk-mmlu3-gsm8k4-v1",
+  "stage": "sft",
+  "wandb_run_id": null,
+  "migration_note": "Migrated from the pre-lineage repository layout."
+}

experiments/think-d12-r11.25/sft/smoltalk-mmlu3-gsm8k4-v1/summary.json

@@ -0,0 +1,6 @@
+{
+  "stage": "sft",
+  "experiment_id": "smoltalk-mmlu3-gsm8k4-v1",
+  "checkpoint_step": 1065,
+  "lineage_inferred_from_checkpoint_metadata": true
+}

experiments/think-d12-r11.25/summary.json

@@ -0,0 +1,5 @@
+{
+  "stage": "base",
+  "experiment_id": "think-d12-r11.25",
+  "checkpoint_step": 2362
+}

experiments/think-d12-r20/config.json

@@ -0,0 +1,49 @@
+{
+  "schema_version": 1,
+  "stage": "base",
+  "experiment_id": "think-d12-r20",
+  "dataset": {
+    "adapter": "parquet_shards",
+    "repo": "jbduran/think-dataset",
+    "revision": "main",
+    "validation_shard": 472,
+    "num_train_shards": 44,
+    "download_workers": 4
+  },
+  "tokenizer": {
+    "mode": "train",
+    "max_chars": 2000000000,
+    "doc_cap": 10000,
+    "vocab_size": 32768
+  },
+  "pretokenize": {
+    "enabled": true,
+    "slack": 1.03,
+    "val_tokens": 20971520,
+    "shard_tokens": 100000000,
+    "tokenizer_threads": 8
+  },
+  "training": {
+    "depth": 12,
+    "scaling_params": 110100912,
+    "target_param_data_ratio": 20.0,
+    "window_pattern": "L",
+    "device_batch_size": 16,
+    "total_batch_size": 524288,
+    "save_every": 500,
+    "eval_every": 250,
+    "eval_tokens": 2097152,
+    "core_metric_every": -1,
+    "sample_every": -1
+  },
+  "artifacts": {
+    "repo": "jbduran/think.nano"
+  },
+  "wandb": {
+    "entity": "jbduran-thinkingmachinesncsu",
+    "project": "think.nano",
+    "name": "think-d12-r20",
+    "group": "think-d12",
+    "tags": ["think-dataset", "d12", "ratio20"]
+  }
+}

experiments/think-d12-r20/run.json

@@ -0,0 +1,6 @@
+{
+  "experiment_id": "think-d12-r20",
+  "stage": "base",
+  "wandb_run_id": null,
+  "migration_note": "Migrated from the pre-lineage repository layout."
+}

experiments/think-d12-r20/summary.json

@@ -0,0 +1,5 @@
+{
+  "stage": "base",
+  "experiment_id": "think-d12-r20",
+  "checkpoint_step": 4200
+}