Add original paper ladder checkpoint and evaluation metadata

README.md

@@ -0,0 +1,60 @@
+---
+language:
+- en
+tags:
+- looped-transformer
+- scaling-laws
+- pytorch
+- base-model
+datasets:
+- HuggingFaceFW/fineweb
+pipeline_tag: text-generation
+---
+
+# loop-2-d10
+
+Loop 2, depth coordinate **d10**, from **How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents**.
+
+This is the original final checkpoint used by the paper's FineWeb scaling ladder. It contains 331,939,840 stored parameters in FP32 (1.328 GB). This is a pretrained base language model, without instruction tuning. The checkpoint preserves the original training artifact and does not contain optimizer state for resuming training.
+
+| Setting | Value |
+| --- | --- |
+| Training corpus | FineWeb |
+| Tokenizer | GPT-2 (`tiktoken.get_encoding("gpt2")`) |
+| Vocabulary | 50,257 tokens, padded to 50,304 model rows |
+| Context length | 2,048 |
+| Width | 1280 |
+| Attention heads | 10 |
+| Model depth mode | `loop` |
+| Configured core repetitions | 2 |
+| Final evaluation repetitions | 2 |
+| Pretraining validation NLL | 3.066159 nats/token |
+
+The depth coordinate is the ladder's scaling coordinate; it need not equal the number of executed Transformer blocks. `result.json` provides the model configuration, parameter counts, training settings, and recorded validation metrics. Validation NLL is measured on the pretraining corpus and is distinct from CORE answer NLL.
+
+## Files
+
+- `final.pt`: original PyTorch checkpoint, including learned weights and training arguments.
+- `result.json`: portable training metadata for reconstructing the model.
+- `SHA256SUMS`: checksum of the original checkpoint.
+
+## Evaluation
+
+Use the paper's codebase at [cue-engineering/loop](https://github.com/cue-engineering/loop), install its dependencies, and download this repository's files:
+
+```python
+from huggingface_hub import snapshot_download
+snapshot_download(repo_id="CharlieChen/loop-2-d10", local_dir="checkpoints/loop-2-d10")
+```
+
+From the codebase root, run a bounded GPU smoke evaluation:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python eval.py \
+  --checkpoint checkpoints/loop-2-d10/final.pt \
+  --result-json checkpoints/loop-2-d10/result.json \
+  --max-per-task 10 --seeds 0 1 2 \
+  --out checkpoints/loop-2-d10/core_smoke.json
+```
+
+The paper uses H100 GPUs, FlashAttention-3, and bfloat16 autocast. The custom `TransformerGPT` model is reconstructed by the paper code; this artifact is not a Transformers `AutoModel` checkpoint. Full CORE evaluation uses all 22 tasks and seeds 0/1/2; omit `--max-per-task` to run it. Smoke scores are not full-suite paper results.

SHA256SUMS

@@ -0,0 +1 @@
+109cbe793c74e24490f7e52a2b9fecb30928a63cfaf87804a82c967f107ad20d  final.pt

final.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:109cbe793c74e24490f7e52a2b9fecb30928a63cfaf87804a82c967f107ad20d
+size 1327779695

result.json

@@ -0,0 +1,110 @@
+{
+  "parameter_count": 331939840,
+  "parameter_summary": {
+    "total": 331939840,
+    "parts": {
+      "blocks": 203161600,
+      "lm_head": 64389120,
+      "other": 64389120
+    }
+  },
+  "n_layer": 10,
+  "n_head": 10,
+  "n_embd": 1280,
+  "head_dim": 128,
+  "depth_scale_mode": "loop",
+  "core_repetitions": 2,
+  "core_layer_count": 4,
+  "resolved_core_layer_count": 4,
+  "prelude_layer_count": 3,
+  "resolved_prelude_layer_count": 3,
+  "coda_layer_count": 3,
+  "resolved_coda_layer_count": 3,
+  "effective_depth": 14,
+  "expected_effective_depth": 14.0,
+  "recurrence_rmsnorm": "recurrence_only",
+  "recurrence_emb_alpha": 0.707106781187,
+  "decoder_inject": true,
+  "core_residual_branch_scale": 1.0,
+  "resolved_core_residual_branch_scale": null,
+  "core_repetition_schedule": "none",
+  "core_repetition_crossover_fraction": null,
+  "core_repetition_low": 1,
+  "core_repetition_breakpoints": null,
+  "lsched_lr_warmup_steps": 0,
+  "lsched_lr_change_steps": [],
+  "dep_grow_events": [],
+  "target_core_repetitions": 2,
+  "final_eval_active_core_repetitions": 2,
+  "observed_train_core_repetitions_mean": null,
+  "observed_train_core_repetitions_min": null,
+  "observed_train_core_repetitions_max": null,
+  "observed_train_core_repetitions_histogram": null,
+  "recurrence_truncation_schedule": "none",
+  "recurrence_bp_window": 0,
+  "matrix_lr": 0.03002410777473192,
+  "embedding_lr": 0.009607714487914214,
+  "unembedding_lr": 0.003396840033058765,
+  "lr_multiplier": 1.0,
+  "effective_matrix_lr": 0.03002410777473192,
+  "effective_embedding_lr": 0.009607714487914214,
+  "effective_unembedding_lr": 0.003396840033058765,
+  "weight_decay": 0.05,
+  "matrix_weight_decay": 0.05,
+  "embedding_weight_decay": 0.05,
+  "unembedding_weight_decay": 0.05,
+  "embedding_wd_skip_norm": 0.0,
+  "adam_betas": [
+    0.8,
+    0.95
+  ],
+  "adam_eps": 1e-10,
+  "lr_schedule": "wsd",
+  "warmup_steps": 40,
+  "warmdown_ratio": 1.0,
+  "output_multiplier": 1.0,
+  "residual_branch_multiplier": 0.25,
+  "wte_init_std": 0.02,
+  "uniform_init_scale": 0.0441941738242,
+  "total_batch_size": 524288,
+  "device_batch_size": 32,
+  "num_epochs": 1,
+  "estimated_num_iterations": 3835,
+  "lr_schedule_steps": 3799,
+  "resolved_lr_schedule_steps": 3799,
+  "max_train_steps": 3799,
+  "resolved_max_train_steps": 3799,
+  "completed_epochs": 0,
+  "seed": 42,
+  "train_token_limit": 2011687813,
+  "val_token_limit": 10000000,
+  "token_limit_cutoff_devices": 8,
+  "active_train_tokens": 2011687813,
+  "packed_train_sequences_per_epoch": 981760,
+  "packed_train_tokens_per_epoch": 2010644480,
+  "active_val_packed_sequences": 4864,
+  "active_val_packed_tokens": 9961472,
+  "val_loss": 3.066159298545436,
+  "final_val_loss": 3.066159298545436,
+  "final_train_loss": 3.0773589121392595,
+  "best_val_loss": 3.066159298545436,
+  "best_val_bpb": 0.9965345135869775,
+  "best_val_epoch": 1.0,
+  "train_val_per_epoch": [
+    {
+      "epoch": 1,
+      "step": 3799,
+      "train_loss": 3.0773589121392595,
+      "val_loss": 3.066159298545436,
+      "val_bpb": 0.9965345135869775
+    }
+  ],
+  "core_block_residual_rms": 1.038166880607605,
+  "last_train_mfu": 44.83289602365357,
+  "last_train_tok_per_sec": 1533450,
+  "last_train_step_time_seconds": 0.3419008255004883,
+  "total_training_time_seconds": 1306.2477164268494,
+  "total_wall_time_seconds": 1356.5711660385132,
+  "sequence_len": 2048,
+  "window_pattern": "L"
+}