Packaging: token IDs, raw checkpoint, nanochat tokenizer

generation_config.json sets only bos/eos/pad, read from this repository's own tokenizer rather than copied from v1. No decoding parameters are prescribed: no tested configuration satisfied the preregistered gate of zero looping and zero token-cap endings across all six repetition-stress prompts, so the repository declines to recommend one. Transformers therefore defaults to greedy, which did not pass the gate either -- it is what prescribing nothing leaves.

Without an eos_token_id, generate() has no stop condition and runs to max_new_tokens on every call; that is the regression this file prevents. Both the assistant-end token and bos are listed, because the model emits either to end a turn.

Also adds the raw training checkpoint and the nanochat tokenizer artifacts, matching v1's layout. No weight file is touched: model.safetensors is unchanged from revision 2250a44c.

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": false,
+  "bos_token_id": 32759,
+  "eos_token_id": [
+    32763,
+    32759
+  ],
+  "pad_token_id": 32763
+}

meta_000365.json

@@ -0,0 +1,63 @@
+{
+  "step": 365,
+  "val_bpb": null,
+  "model_config": {
+    "sequence_len": 4096,
+    "vocab_size": 32768,
+    "n_layer": 24,
+    "n_head": 12,
+    "n_kv_head": 12,
+    "n_embd": 1536,
+    "window_pattern": "SSSL",
+    "rope_base": 100000
+  },
+  "user_config": {
+    "run": "rost-sft-v2-C-145M-dbs8-ga2",
+    "mixture": "smoltalk",
+    "identity_name": "rost",
+    "device_type": "cuda",
+    "compile_mode": "default",
+    "serial_compile_warmup": true,
+    "eager_optimizer": true,
+    "init_source": "base",
+    "model_tag": "d24",
+    "model_step": 11136,
+    "output_model_tag": "sft-v2-C",
+    "load_optimizer": 1,
+    "plan_dir": "/home/devops/sft-v2-launch-v6-145M/plan_C",
+    "plan_origin_root": "/home/devops",
+    "plan_filtered_root": "/home/devops/sft-v2-filtered-v2",
+    "plan_start_step": 1,
+    "plan_diagnostic_run": false,
+    "num_iterations": -1,
+    "max_seq_len": 4096,
+    "device_batch_size": 8,
+    "total_batch_size": 524288,
+    "embedding_lr": 0.3,
+    "unembedding_lr": 0.008,
+    "matrix_lr": 0.02,
+    "init_lr_frac": 0.8,
+    "warmup_ratio": 0.0,
+    "warmdown_ratio": 0.5,
+    "final_lr_frac": 0.0,
+    "grad_clip": 0.0,
+    "eval_every": -1,
+    "eval_tokens": 20971520,
+    "save_every": 100,
+    "eval_device_batch_size": -1,
+    "chatcore_batch_size": -1,
+    "eval_oom_policy": "contain",
+    "chatcore_every": -1,
+    "chatcore_max_cat": -1,
+    "chatcore_max_sample": 24,
+    "dialogue_frac": 0.0,
+    "english_frac": 0.0,
+    "ro_gsm8k_epochs": 0,
+    "agent_traces": null,
+    "agent_epochs": 1,
+    "agent_frac": 0.0,
+    "agent_max_repeats": 1000,
+    "mmlu_epochs": 3,
+    "gsm8k_epochs": 4
+  }
+}

model_000365.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:397dcd5264e2e0dafadd20da3726d8105cafb0becf43ced607d1cd721b0df121
+size 4227935530

tokenizer/token_bytes.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30d116ce754dedb997736932f5f3c09421f9acfa4d69c81b84ad88d8d52776c4
+size 132677

tokenizer/tokenizer.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b7fa31fb8663860349da7d4c86ebef456b495748c1899fd243024b9fcfef7f8
+size 414480