Add task-level fine-tuning-scope catalogue and preserve historical comparison limits

README.md

@@ -710,6 +710,14 @@ configs:
     path: latency-aware/intercept-grab-fast/mean_rtx3090_h1_s123_275ep_v1/demonstrations/conversion/expected/lerobot/meta/episodes/**/*.parquet
   - split: validation
     path: latency-aware/intercept-grab-fast/mean_rtx3090_h1_s123_275ep_v1/demonstrations/conversion/expected/lerobot__val/meta/episodes/**/*.parquet
+- config_name: flappy_fixed_l2_fs4
+  data_files:
+  - split: train
+    path:
+    - latency-aware/flappy/flappy_fixed_l2_fs4/train.parquet
+  - split: val
+    path:
+    - latency-aware/flappy/flappy_fixed_l2_fs4/val.parquet
 dataset_info:
 - config_name: humanoidbench_balance_simple
   features:
@@ -2735,6 +2743,37 @@ dataset_info:
   - name: deterministic
     dtype: bool
   download_size: 1211318389
+- config_name: flappy_fixed_l2_fs4
+  features:
+  - name: action
+    list: float32
+  - name: action_id
+    dtype: int64
+  - name: action_text
+    dtype: string
+  - name: deterministic
+    dtype: bool
+  - name: done
+    dtype: bool
+  - name: episode_idx
+    dtype: int64
+  - name: image_stack
+    list: image
+  - name: latency_ms
+    dtype: float64
+  - name: latency_raw_frames
+    dtype: int64
+  - name: prompt
+    dtype: string
+  - name: reward
+    dtype: float64
+  - name: seed
+    dtype: int64
+  - name: split
+    dtype: string
+  - name: t
+    dtype: int64
+  download_size: 11243675137
 ---
 
 # LAGEN benchmark-datasets
@@ -2746,6 +2785,7 @@ Training data, episode-level evaluations and paper experiment manifests.
 | Visual history | [Visual history](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/visual-history) |
 | Latency in prompt | [Latency in prompt](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/latency-in-prompt) |
 | Latency transfer | [Latency transfer](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/latency-transfer) |
+| VLA fine-tuning scope | [VLA fine-tuning scope](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/vla-finetuning-scope) |
 | Mean vs. profile training | [Mean vs. profile training](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/mean-vs-profile) |
 
 | Resource | Repository |

experiments/README.md

@@ -5,6 +5,7 @@
 | Visual history | Flappy Bird, Demon Attack, Deadly Corridor | 17 selected/source checkpoint entries | [Visual history](visual-history/README.md) |
 | Latency in prompt | Flappy Bird, Demon Attack | 2 selected/source checkpoint entries; Demon links pending | [Latency in prompt](latency-in-prompt/README.md) |
 | Latency transfer | Flappy Bird, Demon Attack | 17 selected/source checkpoint entries | [Latency transfer](latency-transfer/README.md) |
+| VLA fine-tuning scope | Flappy Bird | 7 scope runs; 2 historical trials | [Details](vla-finetuning-scope/README.md) |
 | Mean vs. profile training | Flappy Bird, Deadly Corridor, Ant, InterceptGrabFast | 8 selected/source checkpoint entries | [Mean vs. profile training](mean-vs-profile/README.md) |
 
 Each task page provides settings, fixed asset references and evaluation evidence. Shared assets have one current location.

experiments/vla-finetuning-scope/README.md

@@ -0,0 +1,8 @@
+# VLA fine-tuning scope
+
+Compare trainable model components.
+
+| Task | Settings and assets |
+|---|---|
+| Flappy Bird | [Final scope comparison](flappy/README.md) |
+| Historical LoRA/full | [Separate source records](archive/flappy/README.md) |

experiments/vla-finetuning-scope/archive/flappy/README.md

@@ -0,0 +1,5 @@
+# Historical LoRA/full — Flappy Bird
+
+The retained HF evaluation has 50 episodes: Full mean 18.5180, LoRA mean 3.8620. The old paper has a separate 100-episode aggregate: Full 22.3080, LoRA 3.4820. The original 100-episode source and exact weight binding remain unrecovered. Both records are preserved without substitution.
+
+`manifest.json` identifies the separate model, adapter, action head, data and evaluation sources.

experiments/vla-finetuning-scope/archive/flappy/manifest.json

@@ -0,0 +1,178 @@
+{
+  "experiment": "vla-finetuning-scope",
+  "archived": true,
+  "runs": [
+    {
+      "run": "openvla_flappy_l2_full",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "historical/openvla_flappy_l2_full"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_l2_full"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_l2_full/final_model/pytorch_model.pt"
+        }
+      ],
+      "source_bindings": [
+        {
+          "source_repo": "latency-sensitive-bench/openvla_flappy_l2_full",
+          "source_revision": "2e8e58c6af1ee6f8a3701388b2552a4fa65a041c",
+          "target_repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+          "target_run": "historical/openvla_flappy_l2_full",
+          "artifact_state": "complete_checkpoint_bundle",
+          "framework": "QwenOFT",
+          "run_id": "openvla_flappy_l2_full",
+          "training_step": 2000,
+          "seed": 42,
+          "training_data_root": "/root/latency-sensitive-bench/playground/Datasets/rl_games",
+          "training_data_mix": "flappy_fixed_l2_fs4",
+          "dataset_reference": null,
+          "original_dataset_config": null,
+          "weights": [
+            {
+              "path": "historical/openvla_flappy_l2_full/final_model/pytorch_model.pt",
+              "identity": "f6be353e457ac9cf6dbda6cd2e93692d797016100b69ecea4bcab0ececb9a076",
+              "source_path": "final_model/pytorch_model.pt",
+              "size": 9805972271
+            }
+          ],
+          "paper_locations": [
+            "paper/legacy/figures/sft_comparisons/lora_vs_full"
+          ],
+          "paper_identity_confirmed": false,
+          "interpretation": "The legacy experiment and training models are documented. The paper 100-episode raw evaluation and its exact weight binding were not recovered. The separate HF 50-episode record is preserved unchanged.",
+          "unpublished_training_data": "The recorded local training directory has no recovered exact HF source. No replacement dataset is assumed.",
+          "paper_evaluation_record_difference": [
+            {
+              "method": "full",
+              "hf_episode_count": 50,
+              "hf_mean_return": 18.518000243455173,
+              "paper_episode_count": 100,
+              "paper_mean_return": 22.308000287115576,
+              "same_evaluation_record": false,
+              "hf_timestamp": "2026-06-02T23:20:12.557796+00:00",
+              "note": "The HF 50-episode record and paper 100-episode aggregate are retained separately. The paper local per-episode source files were not recovered."
+            },
+            {
+              "method": "lora",
+              "hf_episode_count": 50,
+              "hf_mean_return": 3.8620000724494457,
+              "paper_episode_count": 100,
+              "paper_mean_return": 3.4820000667870046,
+              "same_evaluation_record": false,
+              "hf_timestamp": "2026-06-02T22:57:10.859128+00:00",
+              "note": "The HF 50-episode record and paper 100-episode aggregate are retained separately. The paper local per-episode source files were not recovered."
+            }
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/openvla_flappy_l2_full"
+      }
+    },
+    {
+      "run": "openvla_flappy_l2_lora",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "historical/openvla_flappy_l2_lora"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_l2_lora"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_l2_lora/final_model/action_model.pt"
+        },
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_l2_lora/final_model/adapter_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "source_repo": "latency-sensitive-bench/openvla_flappy_l2_lora",
+          "source_revision": "427f4c0582d6eb9419fa39f5942db1d59585993c",
+          "target_repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+          "target_run": "historical/openvla_flappy_l2_lora",
+          "artifact_state": "complete_checkpoint_bundle",
+          "framework": "QwenOFT",
+          "run_id": "openvla_flappy_l2_lora",
+          "training_step": 2000,
+          "seed": 42,
+          "training_data_root": "/mnt/data/latency_data/playground/Datasets/rl_games",
+          "training_data_mix": "flappy_fixed_l2_fs4",
+          "dataset_reference": null,
+          "original_dataset_config": null,
+          "weights": [
+            {
+              "path": "historical/openvla_flappy_l2_lora/final_model/action_model.pt",
+              "identity": "e8bc83aaf046d0b6c2045cd38c06875ccc60c808701d6eba10fc491e2b016ffb",
+              "source_path": "final_model/action_model.pt",
+              "size": 131200330
+            },
+            {
+              "path": "historical/openvla_flappy_l2_lora/final_model/adapter_model.safetensors",
+              "identity": "551bd2362cc9548b9d97f3a4f77fcfbe2b28f2725b2897ab42b996485962b8bb",
+              "source_path": "final_model/adapter_model.safetensors",
+              "size": 161192096
+            }
+          ],
+          "paper_locations": [
+            "paper/legacy/figures/sft_comparisons/lora_vs_full"
+          ],
+          "paper_identity_confirmed": false,
+          "interpretation": "The legacy experiment and training models are documented. The paper 100-episode raw evaluation and its exact weight binding were not recovered. The separate HF 50-episode record is preserved unchanged.",
+          "unpublished_training_data": "The recorded local training directory has no recovered exact HF source. No replacement dataset is assumed.",
+          "paper_evaluation_record_difference": [
+            {
+              "method": "full",
+              "hf_episode_count": 50,
+              "hf_mean_return": 18.518000243455173,
+              "paper_episode_count": 100,
+              "paper_mean_return": 22.308000287115576,
+              "same_evaluation_record": false,
+              "hf_timestamp": "2026-06-02T23:20:12.557796+00:00",
+              "note": "The HF 50-episode record and paper 100-episode aggregate are retained separately. The paper local per-episode source files were not recovered."
+            },
+            {
+              "method": "lora",
+              "hf_episode_count": 50,
+              "hf_mean_return": 3.8620000724494457,
+              "paper_episode_count": 100,
+              "paper_mean_return": 3.4820000667870046,
+              "same_evaluation_record": false,
+              "hf_timestamp": "2026-06-02T22:57:10.859128+00:00",
+              "note": "The HF 50-episode record and paper 100-episode aggregate are retained separately. The paper local per-episode source files were not recovered."
+            }
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/openvla_flappy_l2_lora"
+      }
+    }
+  ],
+  "training_data": {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+    "path": "latency-aware/flappy/flappy_fixed_l2_fs4"
+  }
+}

experiments/vla-finetuning-scope/flappy/README.md

@@ -0,0 +1,53 @@
+# VLA fine-tuning scope — Flappy Bird
+
+# VLA fine-tuning scope
+
+This experiment asks how much of OpenVLA must be fine-tuned for latency-aware
+control. The initial hypothesis was that preserving the pretrained visual
+representation and adapting only the upper language-model layers plus the
+action head would be sufficient. The final comparison does not support that
+hypothesis: full-model fine-tuning achieves the highest evaluation return.
+
+The result shows no evidence that full-model fine-tuning harms downstream
+control in this setting. It does **not** establish that the pretrained visual
+representation remains unchanged or that perception itself improves, because
+the experiment contains no perception-specific metric.
+
+## Reported comparison
+
+The figure reports the final checkpoint for each of three fine-tuning scopes:
+
+| Figure label | Trainable scope |
+| --- | --- |
+| `Top-8 + vision` | Vision tower and connector, tied embeddings/language-model head, decoder layers 28--35, and action head |
+| `Top-8 only` | Decoder layers 28--35 and action head; vision tower, connector, and tied embeddings/language-model head are frozen |
+| `Full model` | All model parameters |
+
+Qwen3-VL-4B has 36 decoder layers, so layers 28--35 are the upper eight
+layers. `data.csv` contains only these three final checkpoints. The runs use an
+effective batch size of 128 examples (64 x 2 accumulation for the partial
+scopes and 32 x 4 for full-model tuning). Their module-specific learning rates
+are also matched: `2e-5` base, `1e-5` VLA interface, and `1e-4` action head.
+
+## Setting
+
+- Model: OpenVLA initialized from the Qwen3-VL-4B Bridge checkpoint.
+- Environment: Flappy Bird.
+- Training data: 200 episodes collected with fixed two-tick latency.
+- Input: one observation frame.
+- Training duration: 5,000 optimizer steps.
+- Evaluation latency: fixed at two environment ticks.
+- Evaluation: one final checkpoint per scope, 50 episodes, evaluation seed 42,
+  and a 3,600-step episode limit.
+- Metric: episode return.
+
+The bars show mean return. Error bars are episode-level normal-approximation
+95% confidence intervals,
+`1.96 * std_return / sqrt(50)`. They measure variation across evaluation
+episodes from one checkpoint, not variation across independent training
+seeds.
+
+
+## Assets
+
+`manifest.json` records the seven retained scope runs. Only the three final sources in the paper CSV form the plotted comparison. Shared Flappy L2 data remain in `latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep`.

experiments/vla-finetuning-scope/flappy/manifest.json

@@ -0,0 +1,443 @@
+{
+  "experiment": "vla-finetuning-scope",
+  "task": "flappy",
+  "runs": [
+    {
+      "run": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps/checkpoints/steps_4000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 4000,
+          "training_started_at_utc": "2026-06-24T11:59:06.297366Z",
+          "git_commit": "7b77b87cb564e33377cc2a4f17fc513677849877",
+          "wandb_run_id": "w9uahhsz",
+          "checkpoint": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps/checkpoints/steps_4000_model.safetensors",
+          "checkpoint_identity": "074a513f33ae5b81499478f5b7cd258757d271ce29077b30e299b49f5ba81bdb",
+          "source_path": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps/checkpoints/steps_4000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_vit": false,
+            "freeze_tied_embedding": false,
+            "freeze_llm_layers": []
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": []
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_4k_steps"
+      }
+    },
+    {
+      "run": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps/checkpoints/steps_3000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 3000,
+          "training_started_at_utc": "2026-06-24T09:52:32.813741Z",
+          "git_commit": "68f11e4545a2ad73d3783fd50c14ab886749d0fd",
+          "wandb_run_id": "pllmo179",
+          "checkpoint": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps/checkpoints/steps_3000_model.safetensors",
+          "checkpoint_identity": "dee60fc99cf35e2e6d5908129093c8c66c2cafb62d3b87a236d45acd86522d59",
+          "source_path": "flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps/checkpoints/steps_3000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_vit": false,
+            "freeze_tied_embedding": false,
+            "freeze_llm_layers": []
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": []
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_full_tuning_corrected_bs32_3k_steps"
+      }
+    },
+    {
+      "run": "flappy_fix_latency_2_200ep_last_8_layers_corrected_v2",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_last_8_layers_corrected_v2"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_last_8_layers_corrected_v2"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_last_8_layers_corrected_v2/checkpoints/steps_5000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_last_8_layers_corrected_v2",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 5000,
+          "training_started_at_utc": "2026-06-23T16:43:48.375941Z",
+          "git_commit": "fcb51e2694288bdcf8dc8f558e9973cdaaafb944",
+          "wandb_run_id": "z0dfovtb",
+          "checkpoint": "flappy_fix_latency_2_200ep_last_8_layers_corrected_v2/checkpoints/steps_5000_model.safetensors",
+          "checkpoint_identity": "b2375e35dca20389af04508aba967acffda126a63d3daa8344629d271bdcac3d",
+          "source_path": "flappy_fix_latency_2_200ep_last_8_layers_corrected_v2/checkpoints/steps_5000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_vit": true,
+            "freeze_tied_embedding": true,
+            "freeze_llm_layers": [
+              0,
+              27
+            ]
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": [
+            "exps/vla_finetuning_scope"
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_last_8_layers_corrected_v2"
+      }
+    },
+    {
+      "run": "flappy_fix_latency_2_200ep_last_8_layers_corrected",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_last_8_layers_corrected"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_last_8_layers_corrected"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_last_8_layers_corrected/checkpoints/steps_5000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_last_8_layers_corrected",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 5000,
+          "training_started_at_utc": "2026-06-23T10:13:20.354969Z",
+          "git_commit": "b6f4f971098c0bc5e340ef616b823a4ca769c7e4",
+          "wandb_run_id": "fomfbmpo",
+          "checkpoint": "flappy_fix_latency_2_200ep_last_8_layers_corrected/checkpoints/steps_5000_model.safetensors",
+          "checkpoint_identity": "750786e2eac8b455ab6e8232416d692e8ad1d5b7a56d386b40bf53a9462ba436",
+          "source_path": "flappy_fix_latency_2_200ep_last_8_layers_corrected/checkpoints/steps_5000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_llm_bottom_ratio": 0.7778
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": [
+            "exps/vla_finetuning_scope"
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_last_8_layers_corrected"
+      }
+    },
+    {
+      "run": "flappy_fix_latency_2_200ep_full_tuning_corrected",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_full_tuning_corrected"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning_corrected"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning_corrected/checkpoints/steps_5000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_full_tuning_corrected",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 5000,
+          "training_started_at_utc": "2026-06-23T20:08:09.389236Z",
+          "git_commit": "b6f4f971098c0bc5e340ef616b823a4ca769c7e4",
+          "wandb_run_id": "jd1xm599",
+          "checkpoint": "flappy_fix_latency_2_200ep_full_tuning_corrected/checkpoints/steps_5000_model.safetensors",
+          "checkpoint_identity": "ff343b81e73667493a90c8c0696872ab7f3ed1bf55fae59cf8b57f9785b040eb",
+          "source_path": "flappy_fix_latency_2_200ep_full_tuning_corrected/checkpoints/steps_5000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_llm_bottom_ratio": 0.0
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": [
+            "exps/vla_finetuning_scope"
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_full_tuning_corrected"
+      }
+    },
+    {
+      "run": "flappy_fix_latency_2_200ep_last_8_layers",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_last_8_layers"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_last_8_layers"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_last_8_layers/checkpoints/steps_5000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_last_8_layers",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 5000,
+          "training_started_at_utc": "2026-06-23T03:14:55.765555Z",
+          "git_commit": "daeb076ebed09fb21a1a1b3ce6ab124c64018212",
+          "wandb_run_id": "kbm7xajn",
+          "checkpoint": "flappy_fix_latency_2_200ep_last_8_layers/checkpoints/steps_5000_model.safetensors",
+          "checkpoint_identity": "8afaa32fca8bd02a62fd251a0457e809370c64803c90166b0962241a2089519a",
+          "source_path": "flappy_fix_latency_2_200ep_last_8_layers/checkpoints/steps_5000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_llm_bottom_ratio": 0.7778
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": [
+            "legacy/vla_hyper"
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_last_8_layers"
+      }
+    },
+    {
+      "run": "flappy_fix_latency_2_200ep_full_tuning",
+      "source": {
+        "repo": "latency-sensitive-bench/vla-finetuning-scope-models",
+        "revision": "1ab96924e691cbe8924ac4671466723bdb5ec3af",
+        "path": "flappy_fix_latency_2_200ep_full_tuning"
+      },
+      "metadata": {
+        "repo": "latency-sensitive-bench/benchmark-models",
+        "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+        "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning"
+      },
+      "checkpoints": [
+        {
+          "repo": "latency-sensitive-bench/benchmark-models",
+          "revision": "f99bd515da2a2a364198848b7f3c2dcaaf9566dd",
+          "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_full_tuning/checkpoints/steps_5000_model.safetensors"
+        }
+      ],
+      "source_bindings": [
+        {
+          "run": "flappy_fix_latency_2_200ep_full_tuning",
+          "module": "vla-finetuning-scope",
+          "train_latency_raw_frames": 2,
+          "training_step": 5000,
+          "training_started_at_utc": "2026-06-23T16:41:09.236188Z",
+          "git_commit": "daeb076ebed09fb21a1a1b3ce6ab124c64018212",
+          "wandb_run_id": "hfsfeb3f",
+          "checkpoint": "flappy_fix_latency_2_200ep_full_tuning/checkpoints/steps_5000_model.safetensors",
+          "checkpoint_identity": "2ad30759d93084b8a8c6e03129e3239fac9a644b4c17c638deb76d82aa8610d5",
+          "source_path": "flappy_fix_latency_2_200ep_full_tuning/checkpoints/steps_5000_state/model.safetensors",
+          "source_model_revision": "946358fcf7850d259d33e1b822b80cf5aed15577",
+          "freeze_config": {
+            "freeze_modules": "",
+            "freeze_llm_bottom_ratio": 0.0
+          },
+          "dataset": {
+            "repo": "latency-sensitive-bench/benchmark-datasets",
+            "revision": "e0f4007f504978f93c32557ea4e9b97fdf62fcf5",
+            "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep",
+            "source_repo": "latency-sensitive-bench/flappy_200ep",
+            "source_revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+            "train_rows": 330734,
+            "train_episodes": 180,
+            "original_source": {
+              "repo": "latency-sensitive-bench/flappy_200ep",
+              "revision": "5df5688a961dc2c91090155c719bff4a5b248235",
+              "path": "flappy_fix_latency_2_200ep"
+            }
+          },
+          "paper_consumers": [
+            "legacy/vla_hyper"
+          ]
+        }
+      ],
+      "evaluation": {
+        "repo": "latency-sensitive-bench/benchmark-datasets",
+        "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+        "path": "evaluations/flappy/flappy_fix_latency_2_200ep_full_tuning"
+      }
+    }
+  ],
+  "shared_training_data": {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "58157c572dbd2e0e5036ce197c42db975e56ae4e",
+    "path": "latency-aware/flappy/flappy_200ep/flappy_fix_latency_2_200ep"
+  }
+}