BIue-Jay commited on
Commit
ec6f728
·
verified ·
1 Parent(s): b036974

a14v2: last two checkpoints + eval/deploy files + training records

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/cache_cosmos_reason2.py +60 -0
  2. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/check_reason2_offline.py +26 -0
  3. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/g1_feedback76_h16.py +27 -0
  4. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/groot_chain.py +139 -0
  5. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/h2_body31_hand44_h16.py +24 -0
  6. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/h2_body31_hand44_planner7_h16.py +25 -0
  7. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/h2w3_body31_hand44_h16.py +24 -0
  8. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/hb_aug_v2.py +163 -0
  9. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/hb_launch.py +92 -0
  10. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/hb_offline_fix.py +24 -0
  11. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/prepare_groot_data.py +181 -0
  12. a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/verify_groot_cpu.py +57 -0
  13. a14-unscrew-bulb/v2-b512/step-002000/config.json +90 -0
  14. a14-unscrew-bulb/v2-b512/step-002000/dataset_meta/hb-dataset-receipt.json +47 -0
  15. a14-unscrew-bulb/v2-b512/step-002000/dataset_meta/modality.json +84 -0
  16. a14-unscrew-bulb/v2-b512/step-002000/embodiment_id.json +57 -0
  17. a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/conf.yaml +240 -0
  18. a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/config.yaml +238 -0
  19. a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/dataset_statistics.json +1151 -0
  20. a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/final_model_config.json +58 -0
  21. a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/final_processor_config.json +0 -0
  22. a14-unscrew-bulb/v2-b512/step-002000/loss_curve.csv +401 -0
  23. a14-unscrew-bulb/v2-b512/step-002000/model-00001-of-00002.safetensors +3 -0
  24. a14-unscrew-bulb/v2-b512/step-002000/model-00002-of-00002.safetensors +3 -0
  25. a14-unscrew-bulb/v2-b512/step-002000/model.safetensors.index.json +0 -0
  26. a14-unscrew-bulb/v2-b512/step-002000/processor_config.json +1145 -0
  27. a14-unscrew-bulb/v2-b512/step-002000/protocol.json +87 -0
  28. a14-unscrew-bulb/v2-b512/step-002000/statistics.json +0 -0
  29. a14-unscrew-bulb/v2-b512/step-002000/train.log +0 -0
  30. a14-unscrew-bulb/v2-b512/step-002000/trainer_state.json +1234 -0
  31. a14-unscrew-bulb/v2-b512/step-002000/training_args.bin +3 -0
  32. a14-unscrew-bulb/v2-b512/step-002000/training_plan.json +25 -0
  33. a14-unscrew-bulb/v2-b512/step-002000/wandb_config.json +1 -0
  34. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/cache_cosmos_reason2.py +60 -0
  35. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/check_reason2_offline.py +26 -0
  36. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/g1_feedback76_h16.py +27 -0
  37. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/groot_chain.py +139 -0
  38. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/h2_body31_hand44_h16.py +24 -0
  39. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/h2_body31_hand44_planner7_h16.py +25 -0
  40. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/h2w3_body31_hand44_h16.py +24 -0
  41. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/hb_aug_v2.py +163 -0
  42. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/hb_launch.py +92 -0
  43. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/hb_offline_fix.py +24 -0
  44. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/prepare_groot_data.py +181 -0
  45. a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/verify_groot_cpu.py +57 -0
  46. a14-unscrew-bulb/v2-b512/step-004000/config.json +90 -0
  47. a14-unscrew-bulb/v2-b512/step-004000/dataset_meta/hb-dataset-receipt.json +47 -0
  48. a14-unscrew-bulb/v2-b512/step-004000/dataset_meta/modality.json +84 -0
  49. a14-unscrew-bulb/v2-b512/step-004000/embodiment_id.json +57 -0
  50. a14-unscrew-bulb/v2-b512/step-004000/experiment_cfg/conf.yaml +240 -0
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/cache_cosmos_reason2.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Put the full gated nvidia/Cosmos-Reason2-2B (pinned 9ce19a1) into the shared HF hub cache $G/hf-home/hub.
2
+
3
+ GR00T N1.7 builds its backbone with Qwen3VLForConditionalGeneration.from_pretrained("nvidia/Cosmos-Reason2-2B") and
4
+ its processor with Qwen3VLProcessor.from_pretrained(same name) before loading the GR00T checkpoint, so every node needs
5
+ the whole repo resolvable by name. groot_node.sh runs with HF_HUB_OFFLINE=1, so nodes read this cache and never need
6
+ the token. refs/main is pinned to the verified revision.
7
+
8
+ Usage (token from stdin, kept in-process only):
9
+ $G/envs/groot-py312/bin/python cache_cosmos_reason2.py $G/hf-home/hub < /root/.hb-secrets/hf_token
10
+ """
11
+ import hashlib
12
+ import json
13
+ import os
14
+ import sys
15
+ import time
16
+
17
+ from huggingface_hub import HfApi, snapshot_download
18
+
19
+ REPO = "nvidia/Cosmos-Reason2-2B"
20
+ REV = "9ce19a195e423419c349abfc86fd07178b230561"
21
+ cache = os.path.abspath(sys.argv[1])
22
+ token = sys.stdin.readline().strip()
23
+ if not token.startswith("hf_"):
24
+ raise SystemExit("token format rejected")
25
+ api = HfApi(endpoint="https://huggingface.co", token=token)
26
+ info = api.model_info(REPO, revision=REV, files_metadata=True)
27
+ assert info.sha == REV, info.sha
28
+ t0 = time.monotonic()
29
+ for attempt in range(1, 21):
30
+ try:
31
+ path = snapshot_download(REPO, revision=REV, cache_dir=cache, token=token, max_workers=8)
32
+ break
33
+ except Exception as error: # transient stalls
34
+ if any(k in str(error).lower() for k in ("gated", "401", "403")):
35
+ raise
36
+ print(f"retry {attempt}: {type(error).__name__}: {str(error)[:200]}", flush=True)
37
+ time.sleep(10)
38
+ else:
39
+ raise SystemExit("download did not finish")
40
+ token = None
41
+ count, size = 0, 0
42
+ for s in info.siblings:
43
+ f = os.path.join(path, s.rfilename)
44
+ assert os.path.isfile(f), s.rfilename
45
+ if s.size is not None:
46
+ assert os.path.getsize(f) == s.size, s.rfilename
47
+ if s.lfs is not None:
48
+ h = hashlib.sha256()
49
+ with open(f, "rb") as stream:
50
+ for chunk in iter(lambda: stream.read(1 << 22), b""):
51
+ h.update(chunk)
52
+ assert h.hexdigest() == s.lfs.sha256, s.rfilename
53
+ count += 1
54
+ size += s.size or 0
55
+ refs = os.path.join(cache, "models--" + REPO.replace("/", "--"), "refs")
56
+ os.makedirs(refs, exist_ok=True)
57
+ with open(os.path.join(refs, "main"), "w") as f:
58
+ f.write(REV)
59
+ print("CACHE_OK", json.dumps({"repo": REPO, "revision": REV, "snapshot": path, "files": count, "bytes": size,
60
+ "seconds": round(time.monotonic() - t0, 1)}), flush=True)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/check_reason2_offline.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check that GR00T's Cosmos-Reason2-2B pieces load offline from $G/hf-home exactly as groot_node.sh will.
2
+
3
+ Usage: HF_HOME=$G/hf-home HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 $G/envs/groot-py312/bin/python check_reason2_offline.py
4
+ """
5
+ import json
6
+ import os
7
+ import sys
8
+ from pathlib import Path
9
+
10
+ import torch
11
+
12
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
13
+ import hb_offline_fix # noqa: E402,F401
14
+ from transformers import AutoConfig, Qwen3VLForConditionalGeneration, Qwen3VLProcessor
15
+
16
+ assert os.environ.get("HF_HUB_OFFLINE") == "1", "run with HF_HUB_OFFLINE=1"
17
+ name = "nvidia/Cosmos-Reason2-2B"
18
+ torch.set_num_threads(8)
19
+ proc = Qwen3VLProcessor.from_pretrained(name, trust_remote_code=True)
20
+ cfg = AutoConfig.from_pretrained(name, trust_remote_code=True)
21
+ ids = proc.tokenizer("pick up the trash and drop it in the bin")["input_ids"]
22
+ model = Qwen3VLForConditionalGeneration.from_pretrained(name, torch_dtype=torch.bfloat16)
23
+ n = sum(p.numel() for p in model.parameters())
24
+ print("REASON2_OFFLINE_OK", json.dumps({"processor": type(proc).__name__, "tokens": len(ids),
25
+ "model_type": cfg.model_type, "params_M": round(n / 1e6, 1),
26
+ "llm_layers": len(model.model.language_model.layers)}), flush=True)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/g1_feedback76_h16.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """G1 feedback76 / H16 GR00T modality config (copy of HumanWAM configs/humanwam/groot/feedback76_h16.py @ c76527b).
2
+
3
+ The matching modality.json is generated per dataset from the recorded joint names (prepare_groot_data.py);
4
+ HumanWAM's feedback76_modality.json has right_arm at observation.state[29:36] (left-hand fingers), which is wrong
5
+ for our G1 exports (right arm is [22:29]).
6
+ """
7
+ from gr00t.configs.data.embodiment_configs import register_modality_config
8
+ from gr00t.data.embodiment_tags import EmbodimentTag
9
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
10
+
11
+ feedback76_h16 = {
12
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["ego_view", "chest"]),
13
+ "state": ModalityConfig(
14
+ delta_indices=[0],
15
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "left_hand", "right_hand",
16
+ "projected_gravity"],
17
+ ),
18
+ "action": ModalityConfig(
19
+ delta_indices=list(range(16)),
20
+ modality_keys=["motion_token", "left_hand_joints", "right_hand_joints"],
21
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
22
+ format=ActionFormat.DEFAULT) for _ in range(3)],
23
+ ),
24
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
25
+ }
26
+
27
+ register_modality_config(feedback76_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/groot_chain.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Per-node GR00T relay: per-layout batch-size probe (largest that fits), then 20-epoch production per task.
2
+
3
+ Run on every node (nohup), node rank from the DLC env RANK: python groot_chain.py [task ...]
4
+ Default tasks: b6 a3 a7 a18 (basket postponed). A phase succeeds only if all 4 node launchers exit 0.
5
+
6
+ Batch (user rule: as large as memory allows, 2026-11-10): descending CANDIDATES, 4 steps each on all 32 GPUs; a
7
+ candidate fits when all launchers exit 0 AND the peak reserved memory of every rank (written by hb_launch.py to
8
+ logs/peak-<phase>-rank<r>.json) is <= MAX_FRAC of the GPU. Node 0 alone judges each probe and writes a verdict file
9
+ the other nodes wait for, so all nodes take the same decision. One decision per robot layout (g1 / h2).
10
+ Outputs: $G/groot-runs-bmax/ (the earlier per-GPU-64 runs stay in $G/groot-runs/).
11
+ """
12
+ import fcntl
13
+ import json
14
+ import math
15
+ import os
16
+ import shutil
17
+ import subprocess
18
+ import sys
19
+ import time
20
+ from pathlib import Path
21
+
22
+ H = Path(__file__).resolve().parent
23
+ G = H.parent
24
+ RUNS = G / os.environ.get("HB_GROOT_RUNS_NAME", "groot-runs-bmax")
25
+ NODE = int(os.environ["RANK"])
26
+ TASKS = sys.argv[1:] or ["b6", "a3", "a7", "a18"] # basket postponed (user plan 2026-11-10)
27
+ CANDIDATES = [320, 288, 256, 224, 192, 160, 128, 96, 64, 48, 32]
28
+ MAX_FRAC = 0.92
29
+ EPOCHS = 20
30
+ PORT_BASE = int(os.environ.get("HB_GROOT_PORT_BASE", "32500")) # probe PORT_BASE+i, production PORT_BASE+50+k
31
+ RUNS.mkdir(parents=True, exist_ok=True)
32
+ os.environ["HB_GROOT_RUNS"] = str(RUNS) # read by groot_node.sh
33
+ lock = (RUNS / f"chain-node-{NODE}.lock").open("a")
34
+ fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
35
+
36
+
37
+ def log(msg):
38
+ line = f"{time.strftime('%Y-%m-%d %H:%M:%S')} node{NODE} {msg}"
39
+ print(line, flush=True)
40
+ with (RUNS / f"chain-node-{NODE}.log").open("a") as f:
41
+ f.write(line + "\n")
42
+
43
+
44
+ def run_phase(task, phase, bs, steps, save, port):
45
+ logs = RUNS / task / "logs"
46
+ exits = [logs / f"{phase}-node-{r}.exit" for r in range(4)]
47
+ if all(p.is_file() for p in exits):
48
+ return all(p.read_text().strip() == "0" for p in exits)
49
+ subprocess.run(["bash", str(H / "groot_node.sh"), task, phase, str(bs), str(steps), str(save), str(port)])
50
+ deadline = time.monotonic() + 600
51
+ while not all(p.is_file() for p in exits):
52
+ if time.monotonic() > deadline:
53
+ log(f"{task}/{phase}: timed out waiting for peer exit files")
54
+ return False
55
+ time.sleep(5)
56
+ ok = all(p.read_text().strip() == "0" for p in exits)
57
+ log(f"{task}/{phase}: bs={bs} steps={steps} -> {'OK' if ok else 'FAIL'}")
58
+ return ok
59
+
60
+
61
+ def peak_frac(task, phase):
62
+ """Max over ranks of peak reserved / total; None when no rank reported."""
63
+ fracs = []
64
+ for f in (RUNS / task / "logs").glob(f"peak-{phase}-rank*.json"):
65
+ try:
66
+ d = json.loads(f.read_text())
67
+ fracs.append(d["peak_reserved_gb"] / d["total_gb"])
68
+ except (ValueError, KeyError, OSError):
69
+ continue
70
+ return (max(fracs), len(fracs)) if fracs else (None, 0)
71
+
72
+
73
+ def probe_verdict(task, cand, port):
74
+ phase = f"probe-bs{cand}"
75
+ verdict = RUNS / task / "logs" / f"{phase}.verdict.json"
76
+ ok = run_phase(task, phase, cand, 4, 1000000, port)
77
+ if NODE == 0 and not verdict.is_file():
78
+ time.sleep(10) # let peak files from the other nodes become visible on NAS
79
+ frac, n = peak_frac(task, phase) if ok else (None, 0)
80
+ fits = bool(ok and frac is not None and frac <= MAX_FRAC)
81
+ verdict.write_text(json.dumps({"per_gpu_batch": cand, "ran_ok": ok, "peak_frac": frac, "ranks_reported": n,
82
+ "max_frac": MAX_FRAC, "fits": fits}) + "\n")
83
+ deadline = time.monotonic() + 600
84
+ while not verdict.is_file():
85
+ if time.monotonic() > deadline:
86
+ log(f"{task}/{phase}: no verdict from node 0")
87
+ return False
88
+ time.sleep(5)
89
+ v = json.loads(verdict.read_text())
90
+ log(f"{task}/{phase}: verdict {v}")
91
+ return v["fits"]
92
+
93
+
94
+ def prune_resume_state(out):
95
+ """Keep weights of every checkpoint but resume state (DeepSpeed optimizer shards, scheduler, RNG) only in the latest."""
96
+ ckpts = sorted((p for p in out.glob("checkpoint-*") if p.name.split("-")[-1].isdigit()),
97
+ key=lambda p: int(p.name.split("-")[-1]))
98
+ for c in ckpts[:-1]:
99
+ for p in list(c.glob("global_step*")) + list(c.glob("optimizer*.pt")) + list(c.glob("scheduler.pt")) \
100
+ + list(c.glob("rng_state*.pth")) + list(c.glob("latest")) + list(c.glob("zero_to_fp32.py")):
101
+ if p.is_dir():
102
+ shutil.rmtree(p, ignore_errors=True)
103
+ else:
104
+ p.unlink(missing_ok=True)
105
+ log(f"{c.name}: resume state dropped (weights kept)")
106
+
107
+
108
+ def batch_for(task):
109
+ layout = "h2" if task in ("a17", "a18") else "g1"
110
+ decision = RUNS / f"batch-decision-{layout}.json"
111
+ if decision.is_file():
112
+ return json.loads(decision.read_text())["per_gpu_batch"]
113
+ for i, cand in enumerate(CANDIDATES):
114
+ if probe_verdict(task, cand, PORT_BASE + i + (20 if layout == "h2" else 0)):
115
+ if NODE == 0:
116
+ decision.write_text(json.dumps({"per_gpu_batch": cand, "global_batch": cand * 32, "probe_task": task,
117
+ "layout": layout, "candidates": CANDIDATES,
118
+ "max_frac": MAX_FRAC}) + "\n")
119
+ return cand
120
+ time.sleep(20)
121
+ log(f"{layout}: no batch size fit; stopping")
122
+ sys.exit(1)
123
+
124
+
125
+ for k, task in enumerate(TASKS):
126
+ bs = batch_for(task)
127
+ log(f"{task}: per-GPU batch {bs} (global {bs * 32})")
128
+ receipt = json.loads((G / "groot-data" / task / "hb-dataset-receipt.json").read_text())
129
+ steps = math.ceil(EPOCHS * receipt["frames"] / (bs * 32))
130
+ save = math.ceil(steps / 2)
131
+ if not run_phase(task, "production", bs, steps, save, PORT_BASE + 50 + k):
132
+ log(f"{task}: production failed; stopping chain")
133
+ sys.exit(1)
134
+ if NODE == 0:
135
+ prune_resume_state(RUNS / task / "production")
136
+ (RUNS / task / "production-complete.json").write_text(json.dumps(
137
+ {"task": task, "frames": receipt["frames"], "epochs": EPOCHS, "per_gpu_batch": bs,
138
+ "global_batch": bs * 32, "max_steps": steps, "save_steps": save}) + "\n")
139
+ log("all tasks done")
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/h2_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A18 H2 (Sharpa hands) GR00T modality config, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/h2_body31_hand44_planner7_h16.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 v2 H2 (Sharpa hands) GR00T modality config (H2 training guide hb_training/docs/h2/a14_training_guide.md 4.1):
2
+ cameras head + chest, state78 = 75 joints + projected gravity, action82 = action.body_target (31) + action.hands (44)
3
+ + action.planner (7, the SONIC planner command), horizon 16, absolute targets. Identical to h2_body31_hand44_h16.py
4
+ except the added `planner` action modality."""
5
+ from gr00t.configs.data.embodiment_configs import register_modality_config
6
+ from gr00t.data.embodiment_tags import EmbodimentTag
7
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
8
+
9
+ h2_body31_hand44_planner7_h16 = {
10
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
11
+ "state": ModalityConfig(
12
+ delta_indices=[0],
13
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
14
+ "projected_gravity"],
15
+ ),
16
+ "action": ModalityConfig(
17
+ delta_indices=list(range(16)),
18
+ modality_keys=["body_target", "left_hand_target", "right_hand_target", "planner"],
19
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
20
+ format=ActionFormat.DEFAULT) for _ in range(4)],
21
+ ),
22
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
23
+ }
24
+
25
+ register_modality_config(h2_body31_hand44_planner7_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/h2w3_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 H2 (Sharpa hands) GR00T modality config with THREE cameras head + chest + wrist (user 2026-11-12; the third-person camera is never a model input). Otherwise identical to h2_body31_hand44_h16.py:, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2w3_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest", "wrist"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2w3_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/hb_aug_v2.py ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Humanoid Benchmark data augmentation v2 (project CLAUDE.md), model-agnostic torch implementation.
2
+
3
+ augment(x, rng, obs_mask=None) takes x: float [T, V, 3, H, W] in [0, 1] (T frames, V cameras) and returns the
4
+ augmented tensor. Each call is one training sample: one of the 15 effects is chosen uniformly and its magnitude
5
+ is scaled by 1.5 around identity; the rest stay at x1. Colour is shared across cameras and time; geometry, light
6
+ patches and white balance are drawn per camera and fixed over time; sensor/optics effects (noise, gaussian blur,
7
+ JPEG, motion blur, sharpen, defocus) apply only to observation frames (obs_mask True).
8
+ Pixel-size parameters are defined at 224 px and scaled by H / 224.
9
+ perturb_joints(q, rng) applies the proprio spec: per-joint bias +-2 deg plus N(0, 0.25 deg) clipped at 3 sigma.
10
+ """
11
+ import io
12
+ import math
13
+
14
+ import numpy as np
15
+ import torch
16
+ import torch.nn.functional as F
17
+ from PIL import Image
18
+ from torchvision.transforms import InterpolationMode
19
+ from torchvision.transforms import functional as TF
20
+
21
+ EFFECTS = ("exposure", "contrast", "saturation", "hue", "gamma", "temperature", "rotation_crop", "noise",
22
+ "gaussian_blur", "jpeg", "motion_blur", "sharpen", "defocus", "light_shadow", "wb_channel")
23
+ CENTER = {"exposure": 0.0, "contrast": 1.0, "saturation": 1.0, "hue": 0.0, "gamma": 1.0, "temperature": 0.0}
24
+ RANGE = {"exposure": (-0.22, 0.22), "contrast": (0.88, 1.12), "saturation": (0.76, 1.24), "hue": (-0.015, 0.015),
25
+ "gamma": (0.89, 1.11), "temperature": (-0.066, 0.066)}
26
+ VERSION = "hb-aug-v2-20261109"
27
+
28
+
29
+ def sample(rng, views, height):
30
+ boost = EFFECTS[int(rng.integers(len(EFFECTS)))]
31
+
32
+ def k(name):
33
+ return 1.5 if boost == name else 1.0
34
+
35
+ px = height / 224.0
36
+ p = {"boost": boost}
37
+ for name, (a, b) in RANGE.items():
38
+ c = CENTER[name]
39
+ p[name] = float(rng.uniform(c + k(name) * (a - c), c + k(name) * (b - c)))
40
+ p["geometry"] = []
41
+ for _ in range(views):
42
+ angle = float(rng.uniform(-1.5 * k("rotation_crop"), 1.5 * k("rotation_crop")))
43
+ p["geometry"].append((angle, 1.0 + abs(angle) * 0.015 + float(rng.uniform(0, 0.02 * k("rotation_crop")))))
44
+ p["noise"] = 4.0 / 255.0 * k("noise") if rng.random() < 0.5 else 0.0
45
+ p["gaussian_blur"] = float(rng.uniform(0.25, 0.65)) * k("gaussian_blur") if rng.random() < 0.15 else 0.0
46
+ if rng.random() < 0.35:
47
+ lo = int(round(100 - k("jpeg") * 30))
48
+ hi = int(round(100 - k("jpeg") * 5))
49
+ p["jpeg"] = int(rng.integers(lo, hi + 1))
50
+ else:
51
+ p["jpeg"] = None
52
+ p["motion_blur"] = ((float(rng.uniform(3, 7)) * k("motion_blur") * px, float(rng.uniform(0, 180)))
53
+ if rng.random() < 0.15 else None)
54
+ p["sharpen"] = float(rng.uniform(0.2, 0.6)) * k("sharpen") if rng.random() < 0.15 else 0.0
55
+ p["defocus"] = float(rng.uniform(1, 2)) * k("defocus") * px if rng.random() < 0.10 else 0.0
56
+ if rng.random() < 0.30:
57
+ g = 0.30 * k("light_shadow")
58
+ p["light_shadow"] = [[(float(rng.uniform(0.1, 0.9)), float(rng.uniform(0.1, 0.9)),
59
+ float(rng.uniform(0.15, 0.35)), float(rng.uniform(1 - g, 1 + g)))
60
+ for _ in range(int(rng.integers(1, 3)))] for _ in range(views)]
61
+ else:
62
+ p["light_shadow"] = None
63
+ if rng.random() < 0.30:
64
+ g = 0.10 * k("wb_channel")
65
+ gains = []
66
+ for _ in range(views):
67
+ row = [1.0, 1.0, 1.0]
68
+ row[int(rng.integers(0, 3))] = float(rng.uniform(1 - g, 1 + g))
69
+ gains.append(row)
70
+ p["wb_channel"] = gains
71
+ else:
72
+ p["wb_channel"] = None
73
+ return p
74
+
75
+
76
+ def _conv(x, kernel):
77
+ n, c, h, w = x.shape
78
+ k = kernel.to(x)[None, None].repeat(c, 1, 1, 1)
79
+ pad = kernel.shape[-1] // 2
80
+ return F.conv2d(F.pad(x, (pad, pad, pad, pad), mode="reflect"), k, groups=c)
81
+
82
+
83
+ def motion_kernel(length, angle_deg):
84
+ size = max(3, int(round(length)) | 1)
85
+ k = torch.zeros(size, size)
86
+ c = size // 2
87
+ th = math.radians(angle_deg)
88
+ for s in np.linspace(-(length - 1) / 2, (length - 1) / 2, 8 * size):
89
+ k[int(round(c - s * math.sin(th))), int(round(c + s * math.cos(th)))] = 1.0
90
+ return k / k.sum()
91
+
92
+
93
+ def disk_kernel(r):
94
+ size = 2 * int(math.ceil(r)) + 1
95
+ c = size // 2
96
+ yy, xx = torch.meshgrid(torch.arange(size), torch.arange(size), indexing="ij")
97
+ k = (((yy - c) ** 2 + (xx - c) ** 2) <= (r + 0.5) ** 2).float()
98
+ return k / k.sum()
99
+
100
+
101
+ def apply(x, p, obs_mask=None, noise_seed=0):
102
+ t, v, c, h, w = x.shape
103
+ x = x.clone().clamp(0, 1)
104
+ for vi, (angle, scale) in enumerate(p["geometry"]):
105
+ x[:, vi] = TF.affine(x[:, vi], angle=angle, translate=[0, 0], scale=scale, shear=[0, 0],
106
+ interpolation=InterpolationMode.BILINEAR)
107
+ if p["light_shadow"] is not None:
108
+ yy, xx = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w), indexing="ij")
109
+ for vi, blobs in enumerate(p["light_shadow"]):
110
+ m = torch.ones(h, w)
111
+ for cx, cy, s, g in blobs:
112
+ m = m + (g - 1.0) * torch.exp(-((xx - cx) ** 2 + (yy - cy) ** 2) / (2 * s * s))
113
+ x[:, vi] = (x[:, vi] * m.clamp(min=0).to(x)).clamp(0, 1)
114
+ if p["wb_channel"] is not None:
115
+ for vi, g in enumerate(p["wb_channel"]):
116
+ x[:, vi] = (x[:, vi] * torch.tensor(g, dtype=x.dtype).view(1, 3, 1, 1)).clamp(0, 1)
117
+ y = x.reshape(t * v, c, h, w)
118
+ y = (y * float(2.0 ** p["exposure"])).clamp(0, 1)
119
+ d = p["temperature"]
120
+ y = (y * y.new_tensor([1.0 + d, 1.0, 1.0 - d]).view(1, 3, 1, 1)).clamp(0, 1)
121
+ y = ((y - 0.5) * p["contrast"] + 0.5).clamp(0, 1)
122
+ y = TF.adjust_saturation(y, p["saturation"])
123
+ y = TF.adjust_hue(y, p["hue"])
124
+ y = y.clamp(0, 1).pow(p["gamma"])
125
+ x = y.reshape(t, v, c, h, w)
126
+ mask = torch.ones(t, dtype=torch.bool) if obs_mask is None else torch.as_tensor(obs_mask, dtype=torch.bool)
127
+ if bool(mask.any()):
128
+ o = x[mask].reshape(-1, c, h, w)
129
+ if p["gaussian_blur"] > 0:
130
+ o = TF.gaussian_blur(o, [3, 3], [p["gaussian_blur"]] * 2)
131
+ if p["motion_blur"] is not None:
132
+ o = _conv(o, motion_kernel(*p["motion_blur"]))
133
+ if p["defocus"] > 0:
134
+ o = _conv(o, disk_kernel(p["defocus"]))
135
+ if p["sharpen"] > 0:
136
+ o = o + p["sharpen"] * (o - TF.gaussian_blur(o, [5, 5], [1.0, 1.0]))
137
+ o = o.clamp(0, 1)
138
+ if p["noise"] > 0:
139
+ gen = torch.Generator().manual_seed(int(noise_seed))
140
+ o = (o + torch.randn(o.shape, generator=gen, dtype=o.dtype) * p["noise"]).clamp(0, 1)
141
+ if p["jpeg"] is not None:
142
+ for i in range(o.shape[0]):
143
+ buf = io.BytesIO()
144
+ TF.to_pil_image(o[i]).save(buf, format="JPEG", quality=int(p["jpeg"]))
145
+ buf.seek(0)
146
+ with Image.open(buf) as im:
147
+ o[i] = TF.pil_to_tensor(im.convert("RGB")).to(o.dtype).div(255.0)
148
+ x[mask] = o.reshape(-1, v, c, h, w)
149
+ return x.clamp(0, 1)
150
+
151
+
152
+ def augment(x, rng, obs_mask=None):
153
+ p = sample(rng, x.shape[1], x.shape[-2])
154
+ return apply(x, p, obs_mask, noise_seed=int(rng.integers(0, 2 ** 62))), p
155
+
156
+
157
+ def perturb_joints(q, rng):
158
+ q = np.asarray(q, dtype=np.float64)
159
+ d = q.shape[-1]
160
+ bias = rng.uniform(-math.radians(2.0), math.radians(2.0), size=d)
161
+ sigma = math.radians(0.25)
162
+ noise = np.clip(rng.normal(0.0, sigma, size=q.shape), -3 * sigma, 3 * sigma)
163
+ return (q + bias + noise).astype(np.float32)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/hb_launch.py ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GR00T N1.7 fine-tuning entry with Humanoid Benchmark augmentation v2.
2
+
3
+ Patches Gr00tN1d7Processor in training mode only:
4
+ - images: native random crop / colour jitter / rotation are bypassed (the deterministic eval resize+center-crop is
5
+ used), then hb_aug_v2 is applied jointly to all cameras of the sample;
6
+ - proprio: the body-joint state groups (HB_BODY_STATE_KEYS) get the v2 bias + noise before normalization.
7
+ Then runs the pinned upstream gr00t/experiment/launch_finetune.py unchanged.
8
+ """
9
+ import json
10
+ import os
11
+ import runpy
12
+ import sys
13
+ from pathlib import Path
14
+
15
+ import numpy as np
16
+ import torch
17
+
18
+ HERE = Path(__file__).resolve().parent
19
+ sys.path.insert(0, str(HERE))
20
+ import hb_offline_fix # noqa: E402,F401 (tokenizer load under HF_HUB_OFFLINE=1, see module docstring)
21
+ import hb_aug_v2 as A # noqa: E402
22
+ from gr00t.model.gr00t_n1d7 import processing_gr00t_n1d7 as P # noqa: E402
23
+
24
+ Proc = P.Gr00tN1d7Processor
25
+ BODY_KEYS = [k for k in os.environ.get("HB_BODY_STATE_KEYS", "left_leg,right_leg,waist,left_arm,right_arm").split(",") if k]
26
+ LOG = os.environ.get("HB_AUG_LOG")
27
+ _state = {"rng": None, "count": 0}
28
+
29
+
30
+ def _rng():
31
+ if _state["rng"] is None:
32
+ _state["rng"] = np.random.default_rng([os.getpid(), int.from_bytes(os.urandom(8), "little")])
33
+ return _state["rng"]
34
+
35
+
36
+ _orig_get, _orig_apply, _orig_call = Proc._get_vlm_inputs, Proc._apply_vlm_processing, Proc.__call__
37
+
38
+
39
+ def _get_vlm_inputs(self, image_keys, images, masks, image_transform, language):
40
+ if getattr(self, "training", False):
41
+ self._hb_views = len(image_keys)
42
+ try:
43
+ return _orig_get(self, image_keys, images, masks, self.eval_image_transform, language)
44
+ finally:
45
+ self._hb_views = None
46
+ return _orig_get(self, image_keys, images, masks, image_transform, language)
47
+
48
+
49
+ def _apply_vlm_processing(self, images, language):
50
+ views = getattr(self, "_hb_views", None)
51
+ if views and isinstance(images, torch.Tensor):
52
+ shape = images.shape
53
+ x = images.reshape(-1, views, *shape[1:]).float().div(255.0)
54
+ x, p = A.augment(x, _rng())
55
+ images = (x * 255.0 + 0.5).clamp(0, 255).to(torch.uint8).reshape(shape)
56
+ _state["count"] += 1
57
+ if LOG and _state["count"] in (1, 50):
58
+ with open(f"{LOG}.{os.getpid()}", "a") as f:
59
+ f.write(json.dumps({"count": _state["count"], "views": views, "hw": list(shape[-2:]),
60
+ "boost": p["boost"], "version": A.VERSION}) + "\n")
61
+ return _orig_apply(self, images, language)
62
+
63
+
64
+ def _call(self, messages):
65
+ if getattr(self, "training", False) and BODY_KEYS:
66
+ states = messages[0]["content"].states
67
+ for key in BODY_KEYS:
68
+ if key in states:
69
+ states[key] = A.perturb_joints(states[key], _rng())
70
+ return _orig_call(self, messages)
71
+
72
+
73
+ Proc._get_vlm_inputs = _get_vlm_inputs
74
+ Proc._apply_vlm_processing = _apply_vlm_processing
75
+ Proc.__call__ = _call
76
+
77
+ if os.environ.get("HB_PATCH_ONLY") != "1":
78
+ entry = Path(os.environ["GROOT_SRC"]) / "gr00t/experiment/launch_finetune.py"
79
+ sys.argv = [str(entry)] + sys.argv[1:]
80
+ try:
81
+ runpy.run_path(str(entry), run_name="__main__")
82
+ finally: # batch probe: report this rank's peak reserved memory (groot_chain.py compares it with MAX_FRAC)
83
+ _peak = os.environ.get("HB_PEAK_FILE")
84
+ if _peak and torch.cuda.is_available():
85
+ try:
86
+ _dev = torch.cuda.current_device()
87
+ with open(f"{_peak}-rank{os.environ.get('RANK', '0')}.json", "w") as _f:
88
+ json.dump({"rank": int(os.environ.get("RANK", "0")),
89
+ "peak_reserved_gb": torch.cuda.max_memory_reserved(_dev) / 1e9,
90
+ "total_gb": torch.cuda.get_device_properties(_dev).total_memory / 1e9}, _f)
91
+ except Exception: # noqa: BLE001 - never mask the training result
92
+ pass
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/hb_offline_fix.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Make tokenizer loading work with HF_HUB_OFFLINE=1 on transformers 4.57.x.
2
+
3
+ PreTrainedTokenizerBase._patch_mistral_regex calls huggingface_hub.model_info() for every hub-named tokenizer with a
4
+ vocab > 100k (to detect Mistral base models). With HF_HUB_OFFLINE=1 that raises OfflineModeIsEnabled even though all
5
+ files are in the cache. GR00T's Cosmos-Reason2-2B / Qwen3-VL tokenizer is not a Mistral tokenizer, so when offline the
6
+ check is skipped and the tokenizer is returned unchanged. Online behaviour is untouched.
7
+ """
8
+ import os
9
+
10
+ from transformers.tokenization_utils_base import PreTrainedTokenizerBase as _PTB
11
+
12
+ if not getattr(_PTB, "_hb_offline_fix", False):
13
+ _orig = _PTB._patch_mistral_regex.__func__
14
+
15
+ def _patch_mistral_regex(cls, tokenizer, *args, **kwargs):
16
+ try:
17
+ return _orig(cls, tokenizer, *args, **kwargs)
18
+ except Exception as error:
19
+ if os.environ.get("HF_HUB_OFFLINE") == "1" or type(error).__name__ == "OfflineModeIsEnabled":
20
+ return tokenizer
21
+ raise
22
+
23
+ _PTB._patch_mistral_regex = classmethod(_patch_mistral_regex)
24
+ _PTB._hb_offline_fix = True
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/prepare_groot_data.py ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Build GR00T dataset views (symlinked episodes + rewritten meta + stats) for the G1 tasks.
2
+
3
+ Usage: <groot python> prepare_groot_data.py <task> [<task> ...] (tasks: b6 basket a3 a7)
4
+ B6 always uses the fixed first 100 episodes (0-99) of b6-v5t7-ready-start-20261109.
5
+ """
6
+ import hashlib
7
+ import json
8
+ import os
9
+ import subprocess
10
+ import sys
11
+ from pathlib import Path
12
+
13
+ H = Path(__file__).resolve().parent
14
+ G = H.parent
15
+ P = G.parent
16
+ OUT = G / "groot-data"
17
+ TASKS = {
18
+ "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100),
19
+ "basket": (P / "basket-current-dynamics-100-v21-b0ff522/data", None), # HF basket-transfer-expert-v1 @current-dynamics-100-v21 b0ff522 (old ed77719 retired)
20
+ "basketv2": (P / "basket-trimv2-c4df4d4/data", None), # HF a20-basket-transfer-pre-reach-trim-expert-v2 @c4df4d40 (user 2026-11-14)
21
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
22
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
23
+ "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None), # HF door-pull-expert-v1 @cf3ee9b (A6)
24
+ "b7green": (P / "b7-green-clothes-hamper-v1-ada288c/data", None), # HF b7-green-clothes-hamper-expert-v1 @ada288c, 100 episodes (user 2026-11-13)
25
+ "a7v2": (P / "a7-cabinet-close-v2-aecf070/data", None), # HF a7-cabinet-close-expert-v2 @aecf070, 100 episodes (user 2026-11-14)
26
+ "b8": (P / "b8-table-wipe-v1-3b2e618/data", None), # HF b8-table-wipe-expert-v1 @3b2e618, 100 episodes (user 2026-11-14)
27
+ "b1": (P / "b1-riceball-microwave-v1-36da298/data", None), # HF b1-riceball-microwave-expert-v1 @36da298, all 115 episodes (user 2026-11-14)
28
+ "b7greenv2": (P / "b7-green-clothes-hamper-v2-c12757e/data", None), # HF b7-green-clothes-hamper @c12757e (v2 at the repo root), all 120 episodes (user 2026-11-14)
29
+ "a10v2": (P / "a10-trolley-push-v2-d3abb4c/data", None), # HF a10-trolley-push-expert-v2 @d3abb4c, all 150 episodes (user 2026-11-13)
30
+ "a1": (P / "a1-ground-object-pickup-v1-f47f07e/data", None), # HF a1-ground-object-pickup-expert-v1 @f47f07e (A1)
31
+ "c6t": (P / "c6-test8-06c231d/data", None), # c6 8-episode adaptation test only
32
+ "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None), # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
33
+ "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
34
+ "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
35
+ "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
36
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # HF a17-unscrew-bottle-expert-v1 @407e6e2
37
+ "a18c2": (P / "a18-h2-dataset-audit-8e956a10/dataset", None), # A18, 2 cameras head + chest (user 2026-11-14 18:2x)
38
+ "a17c2": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # A17, 2 cameras head + chest (user 2026-11-14 18:2x)
39
+ "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None), # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
40
+ "a14v2": (P / "a14-unscrew-bulb-v2-8411a6d/data", None), # HF a14-unscrew-bulb-expert-v2 @8411a6da: + action.planner (H2 guide 2026-11-14)
41
+ }
42
+ CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py",
43
+ "a18c2": "h2_body31_hand44_h16.py", "a17c2": "h2_body31_hand44_h16.py",
44
+ "a14v2": "h2_body31_hand44_planner7_h16.py"} # 2026-11-14 H2 guide: action82 incl. the planner command
45
+ # model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
46
+ CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist"),
47
+ "a18c2": ("head", "chest"), "a17c2": ("head", "chest"), "a14v2": ("head", "chest")} # 2026-11-14: 2-camera re-train (user)
48
+ PLANNER = {"a14v2"} # tasks whose action adds action.planner (7) after body_target + hands (H2 guide a14 2.2)
49
+ H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
50
+ "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}
51
+
52
+
53
+ def h2_modality(info, cams=("head", "chest")):
54
+ f = info["features"]
55
+ names = f["observation.state"]["names"]
56
+ assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
57
+ assert f["observation.projected_gravity"]["shape"] == [3]
58
+ state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
59
+ state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
60
+ return {"state": state,
61
+ "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
62
+ "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
63
+ "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
64
+ "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
65
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
66
+ BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
67
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
68
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
69
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
70
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
71
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
72
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
73
+ GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}
74
+
75
+
76
+ def jl(path):
77
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
78
+
79
+
80
+ def wl(path, rows):
81
+ Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))
82
+
83
+
84
+ def build(task):
85
+ src, limit = TASKS[task]
86
+ info = json.loads((src / "meta/info.json").read_text())
87
+ names = info["features"]["observation.state"]["names"]
88
+ if task in CONFIG:
89
+ idx = list(range(31))
90
+ else:
91
+ idx = [names.index(n) for n in BODY]
92
+ for g, (a, b) in GROUPS.items():
93
+ assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
94
+ for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
95
+ ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
96
+ ("observation.images.ego_view", None), ("observation.images.chest", None)):
97
+ assert key in info["features"], (task, key)
98
+ if dim:
99
+ assert info["features"][key]["shape"] == [dim], (task, key)
100
+ episodes = jl(src / "meta/episodes.jsonl")
101
+ stats = jl(src / "meta/episodes_stats.jsonl")
102
+ n = len(episodes) if limit is None else limit
103
+ episodes, stats = episodes[:n], stats[:n]
104
+ assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
105
+ frames = sum(int(e["length"]) for e in episodes)
106
+ dst = OUT / task
107
+ if dst.exists():
108
+ raise SystemExit(f"{dst} exists")
109
+ (dst / "meta").mkdir(parents=True)
110
+ vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
111
+ for e in episodes:
112
+ i = e["episode_index"]
113
+ chunk = i // int(info["chunks_size"])
114
+ rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
115
+ (dst / rel).parent.mkdir(parents=True, exist_ok=True)
116
+ os.symlink(src / rel, dst / rel)
117
+ for k in vkeys:
118
+ relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
119
+ (dst / relv).parent.mkdir(parents=True, exist_ok=True)
120
+ os.symlink(src / relv, dst / relv)
121
+ if task in CONFIG:
122
+ # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
123
+ # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
124
+ keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
125
+ "timestamp", "frame_index", "episode_index", "index", "task_index"}
126
+ keep |= {f"observation.images.{c}" for c in CAMS[task]}
127
+ keep |= {"action.planner"} if task in PLANNER else set()
128
+ info["features"] = {k: v for k, v in info["features"].items() if k in keep}
129
+ vkeys = [k for k in vkeys if k in keep]
130
+ stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
131
+ for s in stats]
132
+ info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
133
+ total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
134
+ (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
135
+ wl(dst / "meta/episodes.jsonl", episodes)
136
+ wl(dst / "meta/episodes_stats.jsonl", stats)
137
+ (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
138
+ if task in CONFIG:
139
+ modality = h2_modality(info, CAMS[task])
140
+ if task in PLANNER:
141
+ assert info["features"]["action.planner"]["shape"] == [7]
142
+ modality["action"]["planner"] = {"original_key": "action.planner", "start": 0, "end": 7}
143
+ else:
144
+ state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
145
+ for g, (a, b) in GROUPS.items()}
146
+ state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
147
+ right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
148
+ projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
149
+ modality = {"state": state,
150
+ "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
151
+ "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
152
+ "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
153
+ "video": {"ego_view": {"original_key": "observation.images.ego_view"},
154
+ "chest": {"original_key": "observation.images.chest"}},
155
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
156
+ (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
157
+ py = sys.executable
158
+ src_g = G / "HumanWAM/third_party/Isaac-GR00T"
159
+ env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
160
+ done = None
161
+ for tag in ("NEW_EMBODIMENT", "new_embodiment"):
162
+ r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
163
+ "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
164
+ capture_output=True, text=True)
165
+ if r.returncode == 0:
166
+ done = tag
167
+ break
168
+ last = r.stdout[-1500:] + r.stderr[-1500:]
169
+ if done is None:
170
+ raise SystemExit(f"stats failed for {task}:\n{last}")
171
+ assert (dst / "meta/stats.json").is_file()
172
+ receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
173
+ state_dim_raw=len(names), body_index=idx, stats_tag=done,
174
+ modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
175
+ stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
176
+ (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
177
+ print("DATASET_OK", json.dumps(receipt)[:400], flush=True)
178
+
179
+
180
+ for t in sys.argv[1:]:
181
+ build(t)
a14-unscrew-bulb/v2-b512/step-002000/code/hb_groot/verify_groot_cpu.py ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check of the GR00T data path for one task view: loader shapes, v2 augmentation, processor patch import."""
2
+ import json
3
+ import os
4
+ import sys
5
+ from pathlib import Path
6
+
7
+ import numpy as np
8
+ import torch
9
+
10
+ H = Path(__file__).resolve().parent
11
+ sys.path.insert(0, str(H))
12
+ os.environ["HB_PATCH_ONLY"] = "1"
13
+ import hb_launch # noqa: E402,F401 (applies the processor patch)
14
+ import hb_aug_v2 as A # noqa: E402
15
+ from gr00t.data.dataset.lerobot_episode_loader import LeRobotEpisodeLoader # noqa: E402
16
+ from gr00t.model.gr00t_n1d7.processing_gr00t_n1d7 import Gr00tN1d7Processor # noqa: E402
17
+
18
+ view = Path(sys.argv[1])
19
+ if view.name == "a14v2": # 2026-11-14 H2 guide: head + chest, action82 = body_target + hands + planner
20
+ import h2_body31_hand44_planner7_h16 as C # noqa: E402
21
+ cfg = C.h2_body31_hand44_planner7_h16
22
+ elif view.name in ("a18c2", "a17c2"): # 2026-11-14: H2 2-camera re-train (head + chest)
23
+ import h2_body31_hand44_h16 as C # noqa: E402
24
+ cfg = next(v for k, v in vars(C).items() if isinstance(v, dict) and "video" in v)
25
+ elif view.name in ("a14", "a18", "a17"): # every H2 task: 3 cameras head + chest + wrist (user 2026-11-12)
26
+ import h2w3_body31_hand44_h16 as C # noqa: E402
27
+ cfg = C.h2w3_body31_hand44_h16
28
+ else:
29
+ import g1_feedback76_h16 as C # noqa: E402
30
+ cfg = C.feedback76_h16
31
+ loader = LeRobotEpisodeLoader(view, cfg)
32
+ df = loader[0]
33
+ cols = list(df.columns)
34
+ row = df.iloc[20]
35
+ out = {"episodes": len(loader), "columns": cols[:40], "rows_ep0": len(df)}
36
+ state_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("state.")}
37
+ action_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("action.")}
38
+ out.update(state_dims=state_dims, action_dims=action_dims, state_total=sum(state_dims.values()),
39
+ action_total=sum(action_dims.values()))
40
+ vids = [c for c in cols if c.startswith("video.")]
41
+ imgs = [np.asarray(row[c]) for c in vids]
42
+ out["video"] = {c: list(i.shape) for c, i in zip(vids, imgs)}
43
+ x = torch.stack([torch.from_numpy(i).permute(2, 0, 1).float().div(255) for i in imgs])[None]
44
+ rng = np.random.default_rng(0)
45
+ boosts = {}
46
+ for _ in range(60):
47
+ y, p = A.augment(x, rng)
48
+ boosts[p["boost"]] = boosts.get(p["boost"], 0) + 1
49
+ assert y.shape == x.shape and torch.isfinite(y).all() and float(y.min()) >= 0 and float(y.max()) <= 1
50
+ out["aug_boost_counts_60"] = boosts
51
+ out["mean_abs_change_last"] = round(float((y - x).abs().mean()), 4)
52
+ q = np.zeros((1, 29))
53
+ dq = A.perturb_joints(q, rng)
54
+ out["proprio_max_abs_deg"] = round(float(np.degrees(np.abs(dq).max())), 3)
55
+ out["patched"] = Gr00tN1d7Processor._get_vlm_inputs.__name__ == "_get_vlm_inputs" and \
56
+ Gr00tN1d7Processor.__call__.__module__ == "hb_launch"
57
+ print("CPU_VERIFY", json.dumps(out), flush=True)
a14-unscrew-bulb/v2-b512/step-002000/config.json ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_horizon": 40,
3
+ "add_pos_embed": true,
4
+ "apply_sincos_state_encoding": false,
5
+ "architectures": [
6
+ "Gr00tN1d7"
7
+ ],
8
+ "attn_dropout": 0.2,
9
+ "attn_implementation": null,
10
+ "backbone_embedding_dim": 2048,
11
+ "backbone_trainable_params_fp32": true,
12
+ "color_jitter_params": {
13
+ "brightness": 0.3,
14
+ "contrast": 0.4,
15
+ "hue": 0.08,
16
+ "saturation": 0.5
17
+ },
18
+ "crop_fraction": 0.95,
19
+ "diffusion_model_cfg": {
20
+ "attention_head_dim": 48,
21
+ "dropout": 0.2,
22
+ "final_dropout": true,
23
+ "interleave_self_attention": true,
24
+ "norm_type": "ada_norm",
25
+ "num_attention_heads": 32,
26
+ "num_layers": 32,
27
+ "output_dim": 1024,
28
+ "positional_embeddings": null
29
+ },
30
+ "dtype": "bfloat16",
31
+ "exclude_state": false,
32
+ "formalize_language": true,
33
+ "hidden_size": 1024,
34
+ "image_crop_size": [
35
+ 230,
36
+ 230
37
+ ],
38
+ "image_target_size": [
39
+ 256,
40
+ 256
41
+ ],
42
+ "letter_box_transform": false,
43
+ "load_bf16": false,
44
+ "max_action_dim": 132,
45
+ "max_num_embodiments": 32,
46
+ "max_seq_len": 1024,
47
+ "max_state_dim": 132,
48
+ "model_dtype": "bfloat16",
49
+ "model_name": "nvidia/Cosmos-Reason2-2B",
50
+ "model_type": "Gr00tN1d7",
51
+ "noise_beta_alpha": 1.5,
52
+ "noise_beta_beta": 1.0,
53
+ "noise_s": 0.999,
54
+ "num_inference_timesteps": 4,
55
+ "num_timestep_buckets": 1000,
56
+ "random_history_crop": true,
57
+ "random_rotation_angle": 0,
58
+ "reproject_vision": false,
59
+ "rtc_ramp_rate": 6.0,
60
+ "select_layer": 16,
61
+ "shortest_image_edge": 256,
62
+ "state_dropout_prob": 0.2,
63
+ "state_gaussian_noise_std": 0.0,
64
+ "strict_action_padding_mask": false,
65
+ "strict_vl_padding_mask": false,
66
+ "transformers_version": "4.57.3",
67
+ "tune_diffusion_model": true,
68
+ "tune_linear": true,
69
+ "tune_llm": true,
70
+ "tune_projector": true,
71
+ "tune_top_llm_layers": 0,
72
+ "tune_visual": true,
73
+ "tune_vlln": true,
74
+ "use_albumentations": true,
75
+ "use_alternate_vl_dit": true,
76
+ "use_flash_attention": true,
77
+ "use_future_tokens": false,
78
+ "use_mean_std": false,
79
+ "use_percentiles": true,
80
+ "use_vl_self_attention": true,
81
+ "use_vlln": true,
82
+ "vl_self_attention_cfg": {
83
+ "attention_head_dim": 64,
84
+ "dropout": 0.2,
85
+ "final_dropout": true,
86
+ "num_attention_heads": 32,
87
+ "num_layers": 4,
88
+ "positional_embeddings": null
89
+ }
90
+ }
a14-unscrew-bulb/v2-b512/step-002000/dataset_meta/hb-dataset-receipt.json ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a14v2",
3
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/a14-unscrew-bulb-v2-8411a6d/data",
4
+ "episodes": 100,
5
+ "frames": 185638,
6
+ "video_keys": [
7
+ "observation.images.head",
8
+ "observation.images.chest"
9
+ ],
10
+ "state_dim_raw": 75,
11
+ "body_index": [
12
+ 0,
13
+ 1,
14
+ 2,
15
+ 3,
16
+ 4,
17
+ 5,
18
+ 6,
19
+ 7,
20
+ 8,
21
+ 9,
22
+ 10,
23
+ 11,
24
+ 12,
25
+ 13,
26
+ 14,
27
+ 15,
28
+ 16,
29
+ 17,
30
+ 18,
31
+ 19,
32
+ 20,
33
+ 21,
34
+ 22,
35
+ 23,
36
+ 24,
37
+ 25,
38
+ 26,
39
+ 27,
40
+ 28,
41
+ 29,
42
+ 30
43
+ ],
44
+ "stats_tag": "NEW_EMBODIMENT",
45
+ "modality_sha256": "4969468a30008415555bd168ad500c64011f3e0a47718724d1318849d4e162fa",
46
+ "stats_sha256": "4be3342282bad2738d3dc1305154d06c47fb31b86e75e1fbd4a29b49c2d211a8"
47
+ }
a14-unscrew-bulb/v2-b512/step-002000/dataset_meta/modality.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "state": {
3
+ "left_leg": {
4
+ "original_key": "observation.state",
5
+ "start": 0,
6
+ "end": 6
7
+ },
8
+ "right_leg": {
9
+ "original_key": "observation.state",
10
+ "start": 6,
11
+ "end": 12
12
+ },
13
+ "waist": {
14
+ "original_key": "observation.state",
15
+ "start": 12,
16
+ "end": 15
17
+ },
18
+ "left_arm": {
19
+ "original_key": "observation.state",
20
+ "start": 15,
21
+ "end": 22
22
+ },
23
+ "right_arm": {
24
+ "original_key": "observation.state",
25
+ "start": 22,
26
+ "end": 29
27
+ },
28
+ "head": {
29
+ "original_key": "observation.state",
30
+ "start": 29,
31
+ "end": 31
32
+ },
33
+ "left_hand": {
34
+ "original_key": "observation.state",
35
+ "start": 31,
36
+ "end": 53
37
+ },
38
+ "right_hand": {
39
+ "original_key": "observation.state",
40
+ "start": 53,
41
+ "end": 75
42
+ },
43
+ "projected_gravity": {
44
+ "original_key": "observation.projected_gravity",
45
+ "start": 0,
46
+ "end": 3
47
+ }
48
+ },
49
+ "action": {
50
+ "body_target": {
51
+ "original_key": "action.body_target",
52
+ "start": 0,
53
+ "end": 31
54
+ },
55
+ "left_hand_target": {
56
+ "original_key": "action.hands",
57
+ "start": 0,
58
+ "end": 22
59
+ },
60
+ "right_hand_target": {
61
+ "original_key": "action.hands",
62
+ "start": 22,
63
+ "end": 44
64
+ },
65
+ "planner": {
66
+ "original_key": "action.planner",
67
+ "start": 0,
68
+ "end": 7
69
+ }
70
+ },
71
+ "video": {
72
+ "head": {
73
+ "original_key": "observation.images.head"
74
+ },
75
+ "chest": {
76
+ "original_key": "observation.images.chest"
77
+ }
78
+ },
79
+ "annotation": {
80
+ "human.task_description": {
81
+ "original_key": "task_index"
82
+ }
83
+ }
84
+ }
a14-unscrew-bulb/v2-b512/step-002000/embodiment_id.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "robocasa_panda_omron": 13,
3
+ "oxe_droid": 17,
4
+ "oxe_fractal": 18,
5
+ "oxe_language_table": 19,
6
+ "oxe_bridge": 20,
7
+ "unknown": 22,
8
+ "gr1_unified": 20,
9
+ "agibot": 26,
10
+ "sim_behavior_r1_pro": 23,
11
+ "xdof": 24,
12
+ "xdof_oss_data": 25,
13
+ "unitree_g1_full_body_with_waist_height_nav_cmd": 25,
14
+ "real_r1_pro_sharpa": 27,
15
+ "real_r1_pro_sharpa_add_view": 27,
16
+ "real_r1_pro_sharpa_relative_arm_joint": 26,
17
+ "real_r1_pro_sharpa_delta_eef": 26,
18
+ "real_r1_pro_sharpa_absolute_eef": 26,
19
+ "real_r1_pro_sharpa_meanstd": 26,
20
+ "real_r1_pro_sharpa_relative_eef": 26,
21
+ "real_r1_pro_sharpa_relative_eef_add_view": 26,
22
+ "real_r1_pro_sharpa_relative_eef_relative_hand": 26,
23
+ "real_r1_pro_sharpa_relative_eef_human": 26,
24
+ "real_r1_pro_sharpa_relative_eef_human_add_view": 26,
25
+ "real_r1_pro_sharpa_relative_eef_human_relative_hand": 26,
26
+ "real_r1_pro_sharpa_relative_eef_egodex": 26,
27
+ "real_r1_pro_sharpa_relative_eef_egodex_relative_hand": 26,
28
+ "real_r1_pro_sharpa_relative_eef_egodex_wrist_only": 26,
29
+ "real_r1_pro_sharpa_relative_eef_maxinsights": 26,
30
+ "real_r1_pro_sharpa_relative_eef_maxinsights_relative_hand": 26,
31
+ "real_r1_pro_sharpa_relative_eef_mecka": 26,
32
+ "real_r1_pro_sharpa_relative_eef_mecka_relative_hand": 26,
33
+ "real_g1_relative_eef_absolute_joints": 25,
34
+ "real_g1_relative_eef_absolute_joints_wrist_cam": 25,
35
+ "real_g1_relative_eef_relative_joints": 25,
36
+ "real_r1_pro_sharpa_relative_eef_relative_hand_relative_joint": 26,
37
+ "real_r1_pro_sharpa_relative_joint": 29,
38
+ "oxe_droid_relative_eef_relative_joint": 24,
39
+ "oxe_droid_relative_eef_relative_joint_swapped": 24,
40
+ "oxe_droid_relative_eef_relative_joint_upweight_z": 24,
41
+ "oxe_droid_relative_eef_relative_joint_upweight_z_swapped": 24,
42
+ "oxe_droid_relative_eef_relative_joint_3view": 24,
43
+ "oxe_droid_relative_eef_relative_joint_3view_swapped": 24,
44
+ "oxe_droid_relative_eef": 24,
45
+ "oxe_droid_joint_position_relative": 24,
46
+ "xdof_relative_eef_relative_joint": 27,
47
+ "xdof_relative_eef_relative_joint_subtask": 27,
48
+ "xdof_relative_eef": 27,
49
+ "xdof_relative_joint": 28,
50
+ "simpler_env_google": 0,
51
+ "simpler_env_widowx": 1,
52
+ "libero_sim": 2,
53
+ "droid_sim": 3,
54
+ "new_embodiment": 10,
55
+ "robocasa_gr1_tabletop": 10,
56
+ "unitree_g1_sonic": 11
57
+ }
a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/conf.yaml ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ load_config_path: null
2
+ model:
3
+ model_type: Gr00tN1d7
4
+ model_dtype: bfloat16
5
+ model_name: nvidia/Cosmos-Reason2-2B
6
+ backbone_model_type: qwen
7
+ model_revision: null
8
+ tune_top_llm_layers: 0
9
+ backbone_embedding_dim: 2048
10
+ tune_llm: true
11
+ tune_visual: true
12
+ select_layer: 12
13
+ reproject_vision: false
14
+ use_flash_attention: true
15
+ load_bf16: false
16
+ backbone_trainable_params_fp32: true
17
+ image_crop_size:
18
+ - 230
19
+ - 230
20
+ image_target_size:
21
+ - 256
22
+ - 256
23
+ shortest_image_edge: null
24
+ crop_fraction: null
25
+ random_rotation_angle: null
26
+ color_jitter_params: null
27
+ use_albumentations_transforms: true
28
+ letter_box_transform: false
29
+ extra_augmentation_config: null
30
+ formalize_language: true
31
+ apply_sincos_state_encoding: false
32
+ use_percentiles: true
33
+ use_relative_action: true
34
+ max_state_dim: 132
35
+ max_action_dim: 132
36
+ action_horizon: 40
37
+ strict_action_padding_mask: null
38
+ hidden_size: 1024
39
+ input_embedding_dim: 1536
40
+ state_history_length: 1
41
+ add_pos_embed: true
42
+ attn_dropout: 0.2
43
+ use_vlln: true
44
+ strict_vl_padding_mask: null
45
+ max_seq_len: 1024
46
+ use_alternate_vl_dit: true
47
+ attend_text_every_n_blocks: 2
48
+ diffusion_model_cfg:
49
+ positional_embeddings: null
50
+ num_layers: 16
51
+ num_attention_heads: 32
52
+ attention_head_dim: 48
53
+ norm_type: ada_norm
54
+ dropout: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
59
+ noise_beta_alpha: 1.5
60
+ noise_beta_beta: 1.0
61
+ noise_s: 0.999
62
+ num_timestep_buckets: 1000
63
+ tune_projector: true
64
+ tune_diffusion_model: true
65
+ tune_vlln: true
66
+ state_dropout_prob: 0.2
67
+ exclude_state: false
68
+ use_mean_std: false
69
+ max_num_embodiments: 32
70
+ data:
71
+ datasets:
72
+ - dataset_paths:
73
+ - /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-data/a14v2
74
+ embodiment_tag: new_embodiment
75
+ mix_ratio: 1.0
76
+ dataset_type: physical_embodiment
77
+ val_dataset_path: null
78
+ modality_configs:
79
+ new_embodiment:
80
+ video:
81
+ delta_indices:
82
+ - 0
83
+ modality_keys:
84
+ - head
85
+ - chest
86
+ sin_cos_embedding_keys: null
87
+ mean_std_embedding_keys: null
88
+ action_configs: null
89
+ state:
90
+ delta_indices:
91
+ - 0
92
+ modality_keys:
93
+ - left_leg
94
+ - right_leg
95
+ - waist
96
+ - left_arm
97
+ - right_arm
98
+ - head
99
+ - left_hand
100
+ - right_hand
101
+ - projected_gravity
102
+ sin_cos_embedding_keys: null
103
+ mean_std_embedding_keys: null
104
+ action_configs: null
105
+ action:
106
+ delta_indices:
107
+ - 0
108
+ - 1
109
+ - 2
110
+ - 3
111
+ - 4
112
+ - 5
113
+ - 6
114
+ - 7
115
+ - 8
116
+ - 9
117
+ - 10
118
+ - 11
119
+ - 12
120
+ - 13
121
+ - 14
122
+ - 15
123
+ modality_keys:
124
+ - body_target
125
+ - left_hand_target
126
+ - right_hand_target
127
+ - planner
128
+ sin_cos_embedding_keys: null
129
+ mean_std_embedding_keys: null
130
+ action_configs:
131
+ - rep: ABSOLUTE
132
+ type: NON_EEF
133
+ format: DEFAULT
134
+ state_key: null
135
+ - rep: ABSOLUTE
136
+ type: NON_EEF
137
+ format: DEFAULT
138
+ state_key: null
139
+ - rep: ABSOLUTE
140
+ type: NON_EEF
141
+ format: DEFAULT
142
+ state_key: null
143
+ - rep: ABSOLUTE
144
+ type: NON_EEF
145
+ format: DEFAULT
146
+ state_key: null
147
+ language:
148
+ delta_indices:
149
+ - 0
150
+ modality_keys:
151
+ - annotation.human.task_description
152
+ sin_cos_embedding_keys: null
153
+ mean_std_embedding_keys: null
154
+ action_configs: null
155
+ download_cache: false
156
+ shard_size: 1024
157
+ episode_sampling_rate: 0.1
158
+ num_shards_per_epoch: 100000
159
+ ds_weights_alpha: null
160
+ override_pretraining_statistics: true
161
+ mode: single_turn
162
+ random_chop: 0.0
163
+ mock_dataset_mode: false
164
+ shuffle: true
165
+ seed: 42
166
+ multiprocessing_context: fork
167
+ allow_padding: false
168
+ subsample_ratio: 1.0
169
+ loader: fast
170
+ fast_loader:
171
+ decode_threads: 2
172
+ ffmpeg_threads: 1
173
+ shuffle_buffer: 0
174
+ min_fill: 0
175
+ image_crop_size:
176
+ - 244
177
+ - 244
178
+ image_target_size:
179
+ - 224
180
+ - 224
181
+ training:
182
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-2x32/a14v2/production
183
+ experiment_name: null
184
+ max_steps: 4000
185
+ global_batch_size: 512
186
+ per_gpu_batch_size: null
187
+ gradient_accumulation_steps: 1
188
+ learning_rate: 0.0001
189
+ lr_scheduler_type: cosine
190
+ weight_decay: 1.0e-05
191
+ warmup_ratio: 0.05
192
+ warmup_steps: 0
193
+ max_grad_norm: 1.0
194
+ optim: adamw_torch
195
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
196
+ skip_weight_loading: false
197
+ tf32: true
198
+ fp16: false
199
+ bf16: true
200
+ eval_bf16: true
201
+ logging_steps: 10
202
+ save_steps: 500
203
+ save_total_limit: 5
204
+ save_vl_model: false
205
+ save_only_model: false
206
+ resume_from_checkpoint: false
207
+ upload_checkpoints: false
208
+ upload_every: 1000
209
+ upload_last_n_checkpoints: 5
210
+ max_concurrent_uploads: 2
211
+ eval_strategy: 'no'
212
+ eval_steps: 500
213
+ eval_set_split_ratio: 0.1
214
+ eval_batch_size: 2
215
+ save_best_eval_metric_name: ''
216
+ save_best_eval_metric_greater_is_better: true
217
+ deepspeed_stage: 2
218
+ gradient_checkpointing: false
219
+ transformers_trust_remote_code: true
220
+ transformers_local_files_only: false
221
+ transformers_cache_dir: null
222
+ transformers_access_token: null
223
+ use_ddp: false
224
+ ddp_bucket_cap_mb: 100
225
+ num_gpus: 16
226
+ dataloader_num_workers: 8
227
+ remove_unused_columns: false
228
+ use_wandb: false
229
+ wandb_project: finetune-gr00t-n1d7
230
+ enable_profiling: false
231
+ max_retries: 3
232
+ assert_loss_less_than: null
233
+ add_rl_callback: false
234
+ enable_open_loop_eval: false
235
+ open_loop_eval_traj_ids:
236
+ - 0
237
+ open_loop_eval_steps_per_traj: 100
238
+ open_loop_eval_plot_indices: null
239
+ max_steps: 4000
240
+ save_steps: 500
a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/config.yaml ADDED
@@ -0,0 +1,238 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ load_config_path: null
2
+ model:
3
+ model_type: Gr00tN1d7
4
+ model_dtype: bfloat16
5
+ model_name: nvidia/Cosmos-Reason2-2B
6
+ backbone_model_type: qwen
7
+ model_revision: null
8
+ tune_top_llm_layers: 0
9
+ backbone_embedding_dim: 2048
10
+ tune_llm: true
11
+ tune_visual: true
12
+ select_layer: 12
13
+ reproject_vision: false
14
+ use_flash_attention: true
15
+ load_bf16: false
16
+ backbone_trainable_params_fp32: true
17
+ image_crop_size:
18
+ - 230
19
+ - 230
20
+ image_target_size:
21
+ - 256
22
+ - 256
23
+ shortest_image_edge: null
24
+ crop_fraction: null
25
+ random_rotation_angle: null
26
+ color_jitter_params: null
27
+ use_albumentations_transforms: true
28
+ letter_box_transform: false
29
+ extra_augmentation_config: null
30
+ formalize_language: true
31
+ apply_sincos_state_encoding: false
32
+ use_percentiles: true
33
+ use_relative_action: true
34
+ max_state_dim: 132
35
+ max_action_dim: 132
36
+ action_horizon: 40
37
+ strict_action_padding_mask: null
38
+ hidden_size: 1024
39
+ input_embedding_dim: 1536
40
+ state_history_length: 1
41
+ add_pos_embed: true
42
+ attn_dropout: 0.2
43
+ use_vlln: true
44
+ strict_vl_padding_mask: null
45
+ max_seq_len: 1024
46
+ use_alternate_vl_dit: true
47
+ attend_text_every_n_blocks: 2
48
+ diffusion_model_cfg:
49
+ positional_embeddings: null
50
+ num_layers: 16
51
+ num_attention_heads: 32
52
+ attention_head_dim: 48
53
+ norm_type: ada_norm
54
+ dropout: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
59
+ noise_beta_alpha: 1.5
60
+ noise_beta_beta: 1.0
61
+ noise_s: 0.999
62
+ num_timestep_buckets: 1000
63
+ tune_projector: true
64
+ tune_diffusion_model: true
65
+ tune_vlln: true
66
+ state_dropout_prob: 0.2
67
+ exclude_state: false
68
+ use_mean_std: false
69
+ max_num_embodiments: 32
70
+ data:
71
+ datasets:
72
+ - dataset_paths:
73
+ - /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-data/a14v2
74
+ embodiment_tag: new_embodiment
75
+ mix_ratio: 1.0
76
+ dataset_type: physical_embodiment
77
+ val_dataset_path: null
78
+ modality_configs:
79
+ new_embodiment:
80
+ video:
81
+ delta_indices:
82
+ - 0
83
+ modality_keys:
84
+ - head
85
+ - chest
86
+ sin_cos_embedding_keys: null
87
+ mean_std_embedding_keys: null
88
+ action_configs: null
89
+ state:
90
+ delta_indices:
91
+ - 0
92
+ modality_keys:
93
+ - left_leg
94
+ - right_leg
95
+ - waist
96
+ - left_arm
97
+ - right_arm
98
+ - head
99
+ - left_hand
100
+ - right_hand
101
+ - projected_gravity
102
+ sin_cos_embedding_keys: null
103
+ mean_std_embedding_keys: null
104
+ action_configs: null
105
+ action:
106
+ delta_indices:
107
+ - 0
108
+ - 1
109
+ - 2
110
+ - 3
111
+ - 4
112
+ - 5
113
+ - 6
114
+ - 7
115
+ - 8
116
+ - 9
117
+ - 10
118
+ - 11
119
+ - 12
120
+ - 13
121
+ - 14
122
+ - 15
123
+ modality_keys:
124
+ - body_target
125
+ - left_hand_target
126
+ - right_hand_target
127
+ - planner
128
+ sin_cos_embedding_keys: null
129
+ mean_std_embedding_keys: null
130
+ action_configs:
131
+ - rep: absolute
132
+ type: non_eef
133
+ format: default
134
+ state_key: null
135
+ - rep: absolute
136
+ type: non_eef
137
+ format: default
138
+ state_key: null
139
+ - rep: absolute
140
+ type: non_eef
141
+ format: default
142
+ state_key: null
143
+ - rep: absolute
144
+ type: non_eef
145
+ format: default
146
+ state_key: null
147
+ language:
148
+ delta_indices:
149
+ - 0
150
+ modality_keys:
151
+ - annotation.human.task_description
152
+ sin_cos_embedding_keys: null
153
+ mean_std_embedding_keys: null
154
+ action_configs: null
155
+ download_cache: false
156
+ shard_size: 1024
157
+ episode_sampling_rate: 0.1
158
+ num_shards_per_epoch: 100000
159
+ ds_weights_alpha: null
160
+ override_pretraining_statistics: true
161
+ mode: single_turn
162
+ random_chop: 0.0
163
+ mock_dataset_mode: false
164
+ shuffle: true
165
+ seed: 42
166
+ multiprocessing_context: fork
167
+ allow_padding: false
168
+ subsample_ratio: 1.0
169
+ loader: fast
170
+ fast_loader:
171
+ decode_threads: 2
172
+ ffmpeg_threads: 1
173
+ shuffle_buffer: 0
174
+ min_fill: 0
175
+ image_crop_size:
176
+ - 244
177
+ - 244
178
+ image_target_size:
179
+ - 224
180
+ - 224
181
+ training:
182
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-2x32/a14v2/production
183
+ experiment_name: null
184
+ max_steps: 4000
185
+ global_batch_size: 512
186
+ per_gpu_batch_size: null
187
+ gradient_accumulation_steps: 1
188
+ learning_rate: 0.0001
189
+ lr_scheduler_type: cosine
190
+ weight_decay: 1.0e-05
191
+ warmup_ratio: 0.05
192
+ warmup_steps: 0
193
+ max_grad_norm: 1.0
194
+ optim: adamw_torch
195
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
196
+ skip_weight_loading: false
197
+ tf32: true
198
+ fp16: false
199
+ bf16: true
200
+ eval_bf16: true
201
+ logging_steps: 10
202
+ save_steps: 500
203
+ save_total_limit: 5
204
+ save_vl_model: false
205
+ save_only_model: false
206
+ resume_from_checkpoint: false
207
+ upload_checkpoints: false
208
+ upload_every: 1000
209
+ upload_last_n_checkpoints: 5
210
+ max_concurrent_uploads: 2
211
+ eval_strategy: 'no'
212
+ eval_steps: 500
213
+ eval_set_split_ratio: 0.1
214
+ eval_batch_size: 2
215
+ save_best_eval_metric_name: ''
216
+ save_best_eval_metric_greater_is_better: true
217
+ deepspeed_stage: 2
218
+ gradient_checkpointing: false
219
+ transformers_trust_remote_code: true
220
+ transformers_local_files_only: false
221
+ transformers_cache_dir: null
222
+ transformers_access_token: null
223
+ use_ddp: false
224
+ ddp_bucket_cap_mb: 100
225
+ num_gpus: 16
226
+ dataloader_num_workers: 8
227
+ remove_unused_columns: false
228
+ use_wandb: false
229
+ wandb_project: finetune-gr00t-n1d7
230
+ enable_profiling: false
231
+ max_retries: 3
232
+ assert_loss_less_than: null
233
+ add_rl_callback: false
234
+ enable_open_loop_eval: false
235
+ open_loop_eval_traj_ids:
236
+ - 0
237
+ open_loop_eval_steps_per_traj: 100
238
+ open_loop_eval_plot_indices: null
a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/dataset_statistics.json ADDED
@@ -0,0 +1,1151 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "state": {
4
+ "left_leg": {
5
+ "min": [
6
+ -0.6128958463668823,
7
+ -0.19324859976768494,
8
+ -0.8298172950744629,
9
+ 0.10861875116825104,
10
+ -0.20572978258132935,
11
+ -1.0465019941329956
12
+ ],
13
+ "max": [
14
+ 0.2436567097902298,
15
+ 0.46225106716156006,
16
+ 0.3263247013092041,
17
+ 1.9302681684494019,
18
+ 0.1985584944486618,
19
+ 0.1742379069328308
20
+ ],
21
+ "mean": [
22
+ -0.08291719108819962,
23
+ 0.04438887536525726,
24
+ -0.01292995736002922,
25
+ 0.28624671697616577,
26
+ -0.018958766013383865,
27
+ -0.180410698056221
28
+ ],
29
+ "std": [
30
+ 0.07462068647146225,
31
+ 0.04940135031938553,
32
+ 0.08979099988937378,
33
+ 0.12407316267490387,
34
+ 0.03175331652164459,
35
+ 0.05499114468693733
36
+ ],
37
+ "q01": [
38
+ -0.3566366705298424,
39
+ -0.10634419076144695,
40
+ -0.30573988318443296,
41
+ 0.16774147838354111,
42
+ -0.1150549115985632,
43
+ -0.39901193052530287
44
+ ],
45
+ "q99": [
46
+ 0.079097608178854,
47
+ 0.20328079134225846,
48
+ 0.1864018699526787,
49
+ 0.9064103084802632,
50
+ 0.08311851464211945,
51
+ -0.031868566237390025
52
+ ]
53
+ },
54
+ "right_leg": {
55
+ "min": [
56
+ -0.814046323299408,
57
+ -0.3444327712059021,
58
+ -0.874907910823822,
59
+ 0.1267164647579193,
60
+ -0.1505868136882782,
61
+ -0.7191075682640076
62
+ ],
63
+ "max": [
64
+ 0.23584064841270447,
65
+ 0.18926510214805603,
66
+ 0.3123853802680969,
67
+ 1.7143502235412598,
68
+ 0.2334698885679245,
69
+ 0.24713154137134552
70
+ ],
71
+ "mean": [
72
+ -0.07489094883203506,
73
+ -0.031710825860500336,
74
+ -0.12531064450740814,
75
+ 0.26502725481987,
76
+ 0.009358185343444347,
77
+ -0.16612687706947327
78
+ ],
79
+ "std": [
80
+ 0.07553238421678543,
81
+ 0.046644601970911026,
82
+ 0.15004462003707886,
83
+ 0.10957752913236618,
84
+ 0.030451182276010513,
85
+ 0.054880429059267044
86
+ ],
87
+ "q01": [
88
+ -0.3749472975730896,
89
+ -0.22074927031993866,
90
+ -0.557280650138855,
91
+ 0.17375476852059366,
92
+ -0.09411303132772446,
93
+ -0.37517826616764066
94
+ ],
95
+ "q99": [
96
+ 0.08031578488647939,
97
+ 0.0876061436533928,
98
+ 0.13781876921653752,
99
+ 0.8577613151073457,
100
+ 0.08570926383137711,
101
+ 0.007561912708916002
102
+ ]
103
+ },
104
+ "waist": {
105
+ "min": [
106
+ -0.15643762052059174,
107
+ -0.39456242322921753,
108
+ -0.04761948063969612
109
+ ],
110
+ "max": [
111
+ 0.2847904860973358,
112
+ 0.1842348724603653,
113
+ 0.5673450231552124
114
+ ],
115
+ "mean": [
116
+ 0.022475101053714752,
117
+ 0.003372820094227791,
118
+ 0.11245314031839371
119
+ ],
120
+ "std": [
121
+ 0.05206042900681496,
122
+ 0.047778401523828506,
123
+ 0.04732291400432587
124
+ ],
125
+ "q01": [
126
+ -0.07448778688907623,
127
+ -0.10977683395147324,
128
+ -0.011013179961591959
129
+ ],
130
+ "q99": [
131
+ 0.1263279996812344,
132
+ 0.15001599639654167,
133
+ 0.20801660135388375
134
+ ]
135
+ },
136
+ "left_arm": {
137
+ "min": [
138
+ -0.6829113960266113,
139
+ -0.3249562680721283,
140
+ -0.8324018716812134,
141
+ -0.9995133280754089,
142
+ -1.9461172819137573,
143
+ -0.4419930875301361,
144
+ -0.6200801730155945
145
+ ],
146
+ "max": [
147
+ 0.7763772010803223,
148
+ 0.570816695690155,
149
+ 1.1427453756332397,
150
+ 1.2946008443832397,
151
+ 0.396441251039505,
152
+ 0.4460962414741516,
153
+ 1.149983286857605
154
+ ],
155
+ "mean": [
156
+ -0.05981207266449928,
157
+ 0.03916569799184799,
158
+ 0.14514833688735962,
159
+ 0.5034250020980835,
160
+ -0.621605396270752,
161
+ 0.035641323775053024,
162
+ 0.09911087155342102
163
+ ],
164
+ "std": [
165
+ 0.2111656814813614,
166
+ 0.1358349621295929,
167
+ 0.3806634545326233,
168
+ 0.594570517539978,
169
+ 0.6502537727355957,
170
+ 0.16242599487304688,
171
+ 0.16147471964359283
172
+ ],
173
+ "q01": [
174
+ -0.5525031477212906,
175
+ -0.24242513820528985,
176
+ -0.46594474285840987,
177
+ -0.6314999204874039,
178
+ -1.6189918386936188,
179
+ -0.3196585050225258,
180
+ -0.19771895661950112
181
+ ],
182
+ "q99": [
183
+ 0.31699694603681594,
184
+ 0.40756282806396504,
185
+ 0.9419007563591004,
186
+ 1.2071343290805816,
187
+ 0.27662204563617715,
188
+ 0.3928562027215959,
189
+ 0.6461342114210162
190
+ ]
191
+ },
192
+ "right_arm": {
193
+ "min": [
194
+ -1.3342407941818237,
195
+ -1.6312137842178345,
196
+ -2.5693843364715576,
197
+ -1.015097975730896,
198
+ -2.498225212097168,
199
+ -0.46122172474861145,
200
+ -1.2372018098831177
201
+ ],
202
+ "max": [
203
+ 1.3257502317428589,
204
+ 0.28701746463775635,
205
+ 1.3928301334381104,
206
+ 1.9761289358139038,
207
+ 1.6500089168548584,
208
+ 0.4516933560371399,
209
+ 1.073415756225586
210
+ ],
211
+ "mean": [
212
+ -0.07064200937747955,
213
+ -0.2800436317920685,
214
+ 0.23499168455600739,
215
+ 0.683804988861084,
216
+ 0.18612521886825562,
217
+ 0.0013837817823514342,
218
+ -0.17755231261253357
219
+ ],
220
+ "std": [
221
+ 0.3420485556125641,
222
+ 0.2882726788520813,
223
+ 0.3892157971858978,
224
+ 0.7240067720413208,
225
+ 0.47399646043777466,
226
+ 0.1443621814250946,
227
+ 0.22828228771686554
228
+ ],
229
+ "q01": [
230
+ -0.9310666924715042,
231
+ -1.1702177929878235,
232
+ -0.5233841043710707,
233
+ -0.9340342110395431,
234
+ -0.3616590654850006,
235
+ -0.4297730204463005,
236
+ -0.8649821281433105
237
+ ],
238
+ "q99": [
239
+ 0.7214082038402557,
240
+ -0.07699784621596331,
241
+ 0.9987939882278447,
242
+ 1.3279817020893097,
243
+ 1.2403862869739535,
244
+ 0.43106062829494485,
245
+ 0.40737970858812333
246
+ ]
247
+ },
248
+ "head": {
249
+ "min": [
250
+ -0.32443222403526306,
251
+ -0.5137076377868652
252
+ ],
253
+ "max": [
254
+ 0.8440645337104797,
255
+ 0.9505477547645569
256
+ ],
257
+ "mean": [
258
+ 0.1802777200937271,
259
+ 0.1982821822166443
260
+ ],
261
+ "std": [
262
+ 0.17640270292758942,
263
+ 0.2836728096008301
264
+ ],
265
+ "q01": [
266
+ -0.2532933488488197,
267
+ -0.35958231806755064
268
+ ],
269
+ "q99": [
270
+ 0.5192202639579773,
271
+ 0.7361457788944246
272
+ ]
273
+ },
274
+ "left_hand": {
275
+ "min": [
276
+ 2.4212948801505263e-07,
277
+ 1.9910926596367062e-07,
278
+ -0.006993828807026148,
279
+ 1.551115076381393e-07,
280
+ -0.04919465631246567,
281
+ -0.0009927804348990321,
282
+ -0.000999723095446825,
283
+ 1.1005165134747585e-07,
284
+ -0.041222814470529556,
285
+ -0.35059306025505066,
286
+ 4.015005217183898e-08,
287
+ 3.39188943598856e-08,
288
+ -0.044367291033267975,
289
+ 2.7533335611451548e-08,
290
+ -0.12385445088148117,
291
+ 6.039768241095089e-09,
292
+ 5.28672128297103e-09,
293
+ 2.0993535798652374e-08,
294
+ 4.511860662859135e-09,
295
+ -0.34986552596092224,
296
+ 3.718351404558007e-09,
297
+ -0.0007463634246960282
298
+ ],
299
+ "max": [
300
+ 0.6663362383842468,
301
+ 0.7547313570976257,
302
+ 0.211256742477417,
303
+ 0.7687531113624573,
304
+ 1.6422386169433594,
305
+ 0.17422638833522797,
306
+ 0.14348424971103668,
307
+ 0.8744160532951355,
308
+ 0.11292567104101181,
309
+ -0.00012538969167508185,
310
+ 1.1353191137313843,
311
+ 1.0877656936645508,
312
+ 0.2749432623386383,
313
+ 0.9350310564041138,
314
+ 0.3845101594924927,
315
+ 0.23132097721099854,
316
+ 0.0563536174595356,
317
+ 0.6766777038574219,
318
+ 0.060973163694143295,
319
+ 0.044674504548311234,
320
+ 0.203311488032341,
321
+ 0.39978837966918945
322
+ ],
323
+ "mean": [
324
+ 0.3482236862182617,
325
+ 0.40993526577949524,
326
+ 0.12880881130695343,
327
+ 0.42788922786712646,
328
+ 1.0061144828796387,
329
+ 0.08961956202983856,
330
+ 0.06922339648008347,
331
+ 0.48949286341667175,
332
+ 0.05179373174905777,
333
+ -0.2140815407037735,
334
+ 0.5811999440193176,
335
+ 0.5792542695999146,
336
+ 0.1312752366065979,
337
+ 0.5052559971809387,
338
+ 0.20629599690437317,
339
+ 0.12213919311761856,
340
+ 0.03359238803386688,
341
+ 0.37536829710006714,
342
+ 0.036314886063337326,
343
+ -0.20700961351394653,
344
+ 0.11387278884649277,
345
+ 0.1411176174879074
346
+ ],
347
+ "std": [
348
+ 0.1662805676460266,
349
+ 0.1923699676990509,
350
+ 0.05474580451846123,
351
+ 0.19785335659980774,
352
+ 0.4380175769329071,
353
+ 0.04487518221139908,
354
+ 0.034851107746362686,
355
+ 0.22133660316467285,
356
+ 0.027133703231811523,
357
+ 0.0880756601691246,
358
+ 0.2917648255825043,
359
+ 0.2829505205154419,
360
+ 0.06694543361663818,
361
+ 0.2412862628698349,
362
+ 0.08273693919181824,
363
+ 0.05835958197712898,
364
+ 0.013655588962137699,
365
+ 0.17167063057422638,
366
+ 0.01462008897215128,
367
+ 0.10084346681833267,
368
+ 0.05005447193980217,
369
+ 0.06493912637233734
370
+ ],
371
+ "q01": [
372
+ 0.00016220964607782662,
373
+ 0.00016220759425777942,
374
+ 1.5003961816546508e-05,
375
+ 0.00016220836550928652,
376
+ 0.031037625130265954,
377
+ -0.0005797499744221568,
378
+ -0.0005871449504047632,
379
+ 0.00016220831181271934,
380
+ -0.0006015802553156391,
381
+ -0.34935444593429565,
382
+ 3.104879579041153e-05,
383
+ 3.104871939285658e-05,
384
+ -0.0005675999401137233,
385
+ 3.1049164572323204e-05,
386
+ -0.02106221681460738,
387
+ 5.599644905487366e-06,
388
+ 5.599406904366333e-06,
389
+ 3.104908533714479e-05,
390
+ 5.599464202532545e-06,
391
+ -0.3491581082344055,
392
+ 5.5994789227042926e-06,
393
+ 6.658128199887869e-07
394
+ ],
395
+ "q99": [
396
+ 0.6099847716093063,
397
+ 0.7054630142450333,
398
+ 0.20482952371239677,
399
+ 0.7329314881563187,
400
+ 1.595677720308304,
401
+ 0.165761756002903,
402
+ 0.1304655821621418,
403
+ 0.8228267818689348,
404
+ 0.10268750801682473,
405
+ -0.07587863780558109,
406
+ 1.058686480522156,
407
+ 1.0297441506385807,
408
+ 0.24550402447581293,
409
+ 0.8799893993139268,
410
+ 0.330337381362915,
411
+ 0.21464521110057833,
412
+ 0.050717052035033704,
413
+ 0.6358237433433533,
414
+ 0.05460452623665333,
415
+ 0.0012227919523138554,
416
+ 0.18715397596359257,
417
+ 0.25647879242897137
418
+ ]
419
+ },
420
+ "right_hand": {
421
+ "min": [
422
+ -0.17600925266742706,
423
+ -0.19547110795974731,
424
+ -0.001285196398384869,
425
+ -0.12617045640945435,
426
+ -0.18237802386283875,
427
+ -0.3570866286754608,
428
+ -0.35529616475105286,
429
+ -0.01663997583091259,
430
+ -0.33883997797966003,
431
+ -0.3577636480331421,
432
+ -0.007328241132199764,
433
+ -0.02236497588455677,
434
+ -0.15625543892383575,
435
+ -0.0035276319831609726,
436
+ -0.35315313935279846,
437
+ -0.008533859625458717,
438
+ -0.004321299493312836,
439
+ -0.0037566684186458588,
440
+ -0.003450079122558236,
441
+ -0.3511973023414612,
442
+ -0.004400239791721106,
443
+ -0.004646706394851208
444
+ ],
445
+ "max": [
446
+ 0.8941553235054016,
447
+ 0.7033463716506958,
448
+ 0.26894673705101013,
449
+ 0.7000588178634644,
450
+ 0.9367015957832336,
451
+ 0.18415626883506775,
452
+ 0.35696694254875183,
453
+ 1.241142749786377,
454
+ 0.32376980781555176,
455
+ -7.032287248875946e-05,
456
+ 1.1047114133834839,
457
+ 1.0755008459091187,
458
+ 0.32843735814094543,
459
+ 1.123056173324585,
460
+ 1.2099076509475708,
461
+ 1.3894870281219482,
462
+ 1.1016167402267456,
463
+ 1.401743769645691,
464
+ 1.0347567796707153,
465
+ 0.18572156131267548,
466
+ 1.3946585655212402,
467
+ 1.7500405311584473
468
+ ],
469
+ "mean": [
470
+ 0.12926290929317474,
471
+ 0.0955263152718544,
472
+ 0.018561597913503647,
473
+ 0.09059379249811172,
474
+ 0.16886106133460999,
475
+ -0.033464472740888596,
476
+ -0.007828655652701855,
477
+ 0.128667950630188,
478
+ 0.02306434139609337,
479
+ -0.1376446634531021,
480
+ 0.14932413399219513,
481
+ 0.19495530426502228,
482
+ 0.0363171249628067,
483
+ 0.20544971525669098,
484
+ 0.16621845960617065,
485
+ 0.2732662260532379,
486
+ 0.17043839395046234,
487
+ 0.23596914112567902,
488
+ 0.19361625611782074,
489
+ -0.04772578924894333,
490
+ 0.2634362578392029,
491
+ 0.3377842307090759
492
+ ],
493
+ "std": [
494
+ 0.2495754212141037,
495
+ 0.19345857203006744,
496
+ 0.04519812390208244,
497
+ 0.1875726580619812,
498
+ 0.27960050106048584,
499
+ 0.08475446701049805,
500
+ 0.0734327957034111,
501
+ 0.2530064880847931,
502
+ 0.08857491612434387,
503
+ 0.08432066440582275,
504
+ 0.2949703633785248,
505
+ 0.3482760488986969,
506
+ 0.08841970562934875,
507
+ 0.36322399973869324,
508
+ 0.33676207065582275,
509
+ 0.48465195298194885,
510
+ 0.30770912766456604,
511
+ 0.42140427231788635,
512
+ 0.342214971780777,
513
+ 0.10488597303628922,
514
+ 0.4642835557460785,
515
+ 0.5944297313690186
516
+ ],
517
+ "q01": [
518
+ -0.0016405727399978775,
519
+ -0.0880109254270792,
520
+ -2.1123528836142215e-06,
521
+ -0.12003309532999992,
522
+ -0.05783637966960668,
523
+ -0.3281420657038689,
524
+ -0.3220851942896843,
525
+ 1.4690435818920378e-05,
526
+ -0.1864601357281208,
527
+ -0.3388528972864151,
528
+ 2.963023574920953e-05,
529
+ 2.9627374242409133e-05,
530
+ -0.06478686399757862,
531
+ 2.962698019473464e-05,
532
+ -0.10028072550892829,
533
+ 5.3524069562627115e-06,
534
+ 5.312631856213557e-06,
535
+ 2.9885372114222264e-05,
536
+ 5.3423715280587205e-06,
537
+ -0.342301287651062,
538
+ 5.351593699742807e-06,
539
+ -8.636927857878618e-05
540
+ ],
541
+ "q99": [
542
+ 0.7435446184873581,
543
+ 0.5788912963867193,
544
+ 0.20758154034614562,
545
+ 0.5714765864610672,
546
+ 0.817389235496521,
547
+ 0.08214768566191212,
548
+ 0.15469318598508838,
549
+ 0.8174264788627625,
550
+ 0.28467825621366505,
551
+ -0.04641571205109352,
552
+ 1.0985167336463928,
553
+ 1.0242142462730408,
554
+ 0.3281265199184418,
555
+ 1.0699113607406616,
556
+ 1.1454014337062837,
557
+ 1.3786902964115142,
558
+ 0.9867417812347412,
559
+ 1.3899610662460327,
560
+ 1.029982089996338,
561
+ 0.08751801565289509,
562
+ 1.383096650838852,
563
+ 1.720791085958481
564
+ ]
565
+ },
566
+ "projected_gravity": {
567
+ "min": [
568
+ -0.6133120656013489,
569
+ -0.17329253256320953,
570
+ -1.0
571
+ ],
572
+ "max": [
573
+ 0.07943684607744217,
574
+ 0.2704276442527771,
575
+ -0.7421032190322876
576
+ ],
577
+ "mean": [
578
+ -0.02038799412548542,
579
+ 0.005426025949418545,
580
+ -0.9988287091255188
581
+ ],
582
+ "std": [
583
+ 0.043265048414468765,
584
+ 0.023551519960165024,
585
+ 0.003424440510575942
586
+ ],
587
+ "q01": [
588
+ -0.161570712774992,
589
+ -0.06989648021757602,
590
+ -0.9999983310699463
591
+ ],
592
+ "q99": [
593
+ 0.03234612852334978,
594
+ 0.07824454523622996,
595
+ -0.9856383764743805
596
+ ]
597
+ }
598
+ },
599
+ "action": {
600
+ "body_target": {
601
+ "min": [
602
+ -1.3360610008239746,
603
+ -0.556527853012085,
604
+ -0.7912365794181824,
605
+ -0.6635185480117798,
606
+ -0.2566780149936676,
607
+ -2.2260026931762695,
608
+ -1.5023866891860962,
609
+ -0.8661556839942932,
610
+ -0.9158138632774353,
611
+ -0.6750026941299438,
612
+ -0.3781881332397461,
613
+ -2.3237345218658447,
614
+ -0.19714340567588806,
615
+ -0.5128103494644165,
616
+ -0.49847859144210815,
617
+ -0.6817602515220642,
618
+ -0.516617476940155,
619
+ -0.8599881529808044,
620
+ -0.9435862302780151,
621
+ -1.9428762197494507,
622
+ -0.4363322854042053,
623
+ -0.34834763407707214,
624
+ -1.4093296527862549,
625
+ -1.6139307022094727,
626
+ -2.552405834197998,
627
+ -0.9861109852790833,
628
+ -2.500661611557007,
629
+ -0.47342196106910706,
630
+ -1.2217304706573486,
631
+ -0.3470655083656311,
632
+ -0.5276036858558655
633
+ ],
634
+ "max": [
635
+ 0.45084860920906067,
636
+ 1.3837697505950928,
637
+ 0.4785388708114624,
638
+ 1.2688502073287964,
639
+ 0.36922043561935425,
640
+ 2.111999988555908,
641
+ 0.71623295545578,
642
+ 0.6330417394638062,
643
+ 0.3554445207118988,
644
+ 1.2778302431106567,
645
+ 0.31986933946609497,
646
+ 2.111999988555908,
647
+ 0.2532210052013397,
648
+ 0.5808194875717163,
649
+ 0.4321470856666565,
650
+ 0.70904141664505,
651
+ 0.6249685883522034,
652
+ 1.1504570245742798,
653
+ 1.307902216911316,
654
+ 0.42155691981315613,
655
+ 0.4363322854042053,
656
+ 1.15058171749115,
657
+ 1.3521455526351929,
658
+ 0.30856841802597046,
659
+ 1.715354084968567,
660
+ 2.028069257736206,
661
+ 1.5654302835464478,
662
+ 0.4363322854042053,
663
+ 1.0768370628356934,
664
+ 0.7877500057220459,
665
+ 0.9628902673721313
666
+ ],
667
+ "mean": [
668
+ -0.0630703940987587,
669
+ 0.04953567311167717,
670
+ -0.015001873485744,
671
+ 0.2614330053329468,
672
+ 0.08157642185688019,
673
+ 0.34592995047569275,
674
+ -0.0006346297450363636,
675
+ -0.04877258464694023,
676
+ -0.1166241243481636,
677
+ 0.28961533308029175,
678
+ -0.07592622190713882,
679
+ 0.4737832844257355,
680
+ 0.020502256229519844,
681
+ 0.018882686272263527,
682
+ -0.03480979800224304,
683
+ -0.05327524244785309,
684
+ 0.034658823162317276,
685
+ 0.1451016366481781,
686
+ 0.4788362979888916,
687
+ -0.6201384663581848,
688
+ 0.0300055593252182,
689
+ 0.10022830218076706,
690
+ -0.05991199240088463,
691
+ -0.3139083683490753,
692
+ 0.2334689348936081,
693
+ 0.650431752204895,
694
+ 0.18299882113933563,
695
+ -0.007910515181720257,
696
+ -0.17922869324684143,
697
+ 0.15655115246772766,
698
+ 0.1988734006881714
699
+ ],
700
+ "std": [
701
+ 0.0934971421957016,
702
+ 0.0962618887424469,
703
+ 0.09199478477239609,
704
+ 0.12340662628412247,
705
+ 0.052925556898117065,
706
+ 0.3717820942401886,
707
+ 0.09755649417638779,
708
+ 0.1034802719950676,
709
+ 0.14349763095378876,
710
+ 0.10886940360069275,
711
+ 0.05236997455358505,
712
+ 0.38055863976478577,
713
+ 0.05160072445869446,
714
+ 0.047542721033096313,
715
+ 0.053849730640649796,
716
+ 0.22080658376216888,
717
+ 0.1632036417722702,
718
+ 0.38015514612197876,
719
+ 0.5764653086662292,
720
+ 0.6518264412879944,
721
+ 0.16254977881908417,
722
+ 0.16356903314590454,
723
+ 0.3674522340297699,
724
+ 0.2754622995853424,
725
+ 0.39030101895332336,
726
+ 0.7063868641853333,
727
+ 0.4758155047893524,
728
+ 0.15067079663276672,
729
+ 0.23079954087734222,
730
+ 0.1638229489326477,
731
+ 0.28449177742004395
732
+ ],
733
+ "q01": [
734
+ -0.40397412568330765,
735
+ -0.1714669668674469,
736
+ -0.3301435789465904,
737
+ 0.026651987489312887,
738
+ -0.0558497430384159,
739
+ -0.7908401769399642,
740
+ -0.41473329067230225,
741
+ -0.38748405933380126,
742
+ -0.5243088698387146,
743
+ 0.029046748317778117,
744
+ -0.26402541756629944,
745
+ -0.7876112502813339,
746
+ -0.07557526983320713,
747
+ -0.11622332006692886,
748
+ -0.12249280296266078,
749
+ -0.5531135147809982,
750
+ -0.37285903096199036,
751
+ -0.46494767397642134,
752
+ -0.6245145499706268,
753
+ -1.620276402235031,
754
+ -0.3199811279773712,
755
+ -0.19791455566883087,
756
+ -0.9092827749252319,
757
+ -1.1615830087661743,
758
+ -0.4696098971366882,
759
+ -0.9515507674217224,
760
+ -0.36861153453588485,
761
+ -0.41725403636693953,
762
+ -0.877388459444046,
763
+ -0.25041723012924194,
764
+ -0.36126713186502457
765
+ ],
766
+ "q99": [
767
+ 0.1537126497924328,
768
+ 0.3397348135709763,
769
+ 0.19410766497254373,
770
+ 0.8858394026756287,
771
+ 0.24553038716316222,
772
+ 1.3337094509601606,
773
+ 0.21554350838065156,
774
+ 0.1667062962055207,
775
+ 0.1400223323702813,
776
+ 0.8292012727260596,
777
+ 0.05381642878055577,
778
+ 1.3321343517303468,
779
+ 0.11868055321276194,
780
+ 0.11623886950314048,
781
+ 0.1212270621210339,
782
+ 0.3132640290260318,
783
+ 0.39469442903995516,
784
+ 0.9415042799711227,
785
+ 1.176945334672928,
786
+ 0.28002428293228154,
787
+ 0.3785510921478272,
788
+ 0.6569059461355217,
789
+ 0.7154888731241231,
790
+ -0.02325495898723599,
791
+ 0.9989206922054292,
792
+ 1.2999485838413238,
793
+ 1.2396190488338472,
794
+ 0.4363322854042053,
795
+ 0.42578777909278914,
796
+ 0.47424808174371724,
797
+ 0.7383069062232972
798
+ ]
799
+ },
800
+ "left_hand_target": {
801
+ "min": [
802
+ 0.0,
803
+ 0.0,
804
+ 0.0,
805
+ 0.0,
806
+ 0.0,
807
+ 0.0,
808
+ 0.0,
809
+ 0.0,
810
+ 0.0,
811
+ -0.34934768080711365,
812
+ 0.0,
813
+ 0.0,
814
+ 0.0,
815
+ 0.0,
816
+ 0.0,
817
+ 0.0,
818
+ 0.0,
819
+ 0.0,
820
+ 0.0,
821
+ -0.34914669394493103,
822
+ 0.0,
823
+ 0.0
824
+ ],
825
+ "max": [
826
+ 0.7311898469924927,
827
+ 0.8207998871803284,
828
+ 0.24490413069725037,
829
+ 0.8429157733917236,
830
+ 1.8877547979354858,
831
+ 0.1810334026813507,
832
+ 0.13143807649612427,
833
+ 0.949565589427948,
834
+ 0.0950046256184578,
835
+ 0.0,
836
+ 1.1989308595657349,
837
+ 1.1410846710205078,
838
+ 0.24130959808826447,
839
+ 0.9818798303604126,
840
+ 0.4285295009613037,
841
+ 0.253095805644989,
842
+ 0.0660354495048523,
843
+ 0.721745491027832,
844
+ 0.06885168701410294,
845
+ 0.0,
846
+ 0.21433521807193756,
847
+ 0.4040643274784088
848
+ ],
849
+ "mean": [
850
+ 0.3615240454673767,
851
+ 0.42453533411026,
852
+ 0.13323697447776794,
853
+ 0.44148048758506775,
854
+ 1.0234434604644775,
855
+ 0.08936263620853424,
856
+ 0.06758644431829453,
857
+ 0.5037830471992493,
858
+ 0.0487588495016098,
859
+ -0.1954365372657776,
860
+ 0.5915658473968506,
861
+ 0.5899481773376465,
862
+ 0.12663695216178894,
863
+ 0.5142788887023926,
864
+ 0.2320898026227951,
865
+ 0.1261432021856308,
866
+ 0.03618995472788811,
867
+ 0.38404759764671326,
868
+ 0.03876037895679474,
869
+ -0.20527827739715576,
870
+ 0.1173388883471489,
871
+ 0.14899969100952148
872
+ ],
873
+ "std": [
874
+ 0.18413126468658447,
875
+ 0.21056896448135376,
876
+ 0.06108909845352173,
877
+ 0.21453315019607544,
878
+ 0.4767066538333893,
879
+ 0.04557064548134804,
880
+ 0.033536482602357864,
881
+ 0.23709090054035187,
882
+ 0.02397112362086773,
883
+ 0.07854247838258743,
884
+ 0.30176860094070435,
885
+ 0.29255473613739014,
886
+ 0.06068819761276245,
887
+ 0.2502109110355377,
888
+ 0.10672258585691452,
889
+ 0.06286828219890594,
890
+ 0.016281168907880783,
891
+ 0.18021008372306824,
892
+ 0.017049144953489304,
893
+ 0.09523232281208038,
894
+ 0.05349619314074516,
895
+ 0.07059493660926819
896
+ ],
897
+ "q01": [
898
+ 0.0,
899
+ 0.0,
900
+ 0.0,
901
+ 0.0,
902
+ 0.0,
903
+ 0.0,
904
+ 0.0,
905
+ 0.0,
906
+ 0.0,
907
+ -0.3087283968925476,
908
+ 0.0,
909
+ 0.0,
910
+ 0.0,
911
+ 0.0,
912
+ 0.0,
913
+ 0.0,
914
+ 0.0,
915
+ 0.0,
916
+ 0.0,
917
+ -0.3490999937057495,
918
+ 0.0,
919
+ 0.0
920
+ ],
921
+ "q99": [
922
+ 0.679345965385437,
923
+ 0.7722799181938171,
924
+ 0.2282198667526245,
925
+ 0.7869092226028442,
926
+ 1.737125277519226,
927
+ 0.16814137995243073,
928
+ 0.12367808073759079,
929
+ 0.8740845918655396,
930
+ 0.08822762221097946,
931
+ 0.0,
932
+ 1.111319899559021,
933
+ 1.073724627494812,
934
+ 0.22454890608787537,
935
+ 0.9176412224769592,
936
+ 0.3930431604385376,
937
+ 0.2343997061252594,
938
+ 0.06176067143678665,
939
+ 0.6640158891677856,
940
+ 0.06512759625911713,
941
+ 0.0,
942
+ 0.19882987439632416,
943
+ 0.26284244656562805
944
+ ]
945
+ },
946
+ "right_hand_target": {
947
+ "min": [
948
+ -0.01080000028014183,
949
+ -0.09933027625083923,
950
+ 0.0,
951
+ -0.1306000053882599,
952
+ 0.0,
953
+ -0.32910001277923584,
954
+ -0.32909998297691345,
955
+ 0.0,
956
+ -0.32909998297691345,
957
+ -0.32910001277923584,
958
+ 0.0,
959
+ 0.0,
960
+ -0.21770872175693512,
961
+ 0.0,
962
+ 0.0,
963
+ 0.0,
964
+ 0.0,
965
+ 0.0,
966
+ 0.0,
967
+ -0.32909998297691345,
968
+ 0.0,
969
+ 0.0
970
+ ],
971
+ "max": [
972
+ 0.8928331732749939,
973
+ 0.7036266922950745,
974
+ 0.241799995303154,
975
+ 0.6738444566726685,
976
+ 1.016309142112732,
977
+ 0.19733072817325592,
978
+ 0.21058975160121918,
979
+ 0.9845873713493347,
980
+ 0.32909998297691345,
981
+ 0.0,
982
+ 1.0911272764205933,
983
+ 1.1062407493591309,
984
+ 0.32910001277923584,
985
+ 1.0623090267181396,
986
+ 1.2097219228744507,
987
+ 1.3762999773025513,
988
+ 1.1076380014419556,
989
+ 1.3830384016036987,
990
+ 1.0229496955871582,
991
+ 0.01119999960064888,
992
+ 1.3762999773025513,
993
+ 1.7253000736236572
994
+ ],
995
+ "mean": [
996
+ 0.12750159204006195,
997
+ 0.09522440284490585,
998
+ 0.021105997264385223,
999
+ 0.08896595984697342,
1000
+ 0.15417326986789703,
1001
+ -0.03003866598010063,
1002
+ -0.0053154537454247475,
1003
+ 0.1288803517818451,
1004
+ 0.02570687048137188,
1005
+ -0.06401088088750839,
1006
+ 0.14916358888149261,
1007
+ 0.19580388069152832,
1008
+ 0.03314713016152382,
1009
+ 0.20606353878974915,
1010
+ 0.1691204011440277,
1011
+ 0.2725496292114258,
1012
+ 0.17041674256324768,
1013
+ 0.23602993786334991,
1014
+ 0.19342400133609772,
1015
+ -0.04756305739283562,
1016
+ 0.2625974416732788,
1017
+ 0.3352619707584381
1018
+ ],
1019
+ "std": [
1020
+ 0.2505542039871216,
1021
+ 0.19649994373321533,
1022
+ 0.05446814000606537,
1023
+ 0.18861454725265503,
1024
+ 0.28459763526916504,
1025
+ 0.08600185066461563,
1026
+ 0.07535401731729507,
1027
+ 0.25728726387023926,
1028
+ 0.09547928720712662,
1029
+ 0.11592736095190048,
1030
+ 0.29567912220954895,
1031
+ 0.35105255246162415,
1032
+ 0.09010125696659088,
1033
+ 0.365320086479187,
1034
+ 0.3332426846027374,
1035
+ 0.4847574532032013,
1036
+ 0.3084282875061035,
1037
+ 0.42318078875541687,
1038
+ 0.3427359461784363,
1039
+ 0.09942680597305298,
1040
+ 0.4633129835128784,
1041
+ 0.5938907861709595
1042
+ ],
1043
+ "q01": [
1044
+ -0.01080000028014183,
1045
+ -0.09929999709129333,
1046
+ 0.0,
1047
+ -0.1306000053882599,
1048
+ 0.0,
1049
+ -0.32910001277923584,
1050
+ -0.32909998297691345,
1051
+ 0.0,
1052
+ -0.20482102036476135,
1053
+ -0.32910001277923584,
1054
+ 0.0,
1055
+ 0.0,
1056
+ -0.11947768211364745,
1057
+ 0.0,
1058
+ 0.0,
1059
+ 0.0,
1060
+ 0.0,
1061
+ 0.0,
1062
+ 0.0,
1063
+ -0.32909998297691345,
1064
+ 0.0,
1065
+ 0.0
1066
+ ],
1067
+ "q99": [
1068
+ 0.7445388436317444,
1069
+ 0.5747254490852356,
1070
+ 0.22886532545089722,
1071
+ 0.5663459300994873,
1072
+ 0.8356737494468689,
1073
+ 0.11156169772148133,
1074
+ 0.1607861965894699,
1075
+ 0.834538459777832,
1076
+ 0.31079334020614624,
1077
+ 0.0,
1078
+ 1.0910999774932861,
1079
+ 1.0164999961853027,
1080
+ 0.32910001277923584,
1081
+ 1.0622999668121338,
1082
+ 1.1383999586105347,
1083
+ 1.3762999773025513,
1084
+ 0.9796000123023987,
1085
+ 1.3830000162124634,
1086
+ 1.0228999853134155,
1087
+ 0.01119999960064888,
1088
+ 1.3762999773025513,
1089
+ 1.7253000736236572
1090
+ ]
1091
+ },
1092
+ "planner": {
1093
+ "min": [
1094
+ 0.0,
1095
+ 0.0,
1096
+ -0.07298091053962708,
1097
+ -1.0,
1098
+ 0.6693262457847595,
1099
+ -0.467713326215744,
1100
+ 0.0
1101
+ ],
1102
+ "max": [
1103
+ 1.0,
1104
+ 1.0,
1105
+ 1.0,
1106
+ 0.15072773396968842,
1107
+ 1.0,
1108
+ 0.7429686188697815,
1109
+ 0.3499999940395355
1110
+ ],
1111
+ "mean": [
1112
+ 0.2536495625972748,
1113
+ 0.167885884642601,
1114
+ 0.9958081841468811,
1115
+ -0.017255336046218872,
1116
+ 0.9955130815505981,
1117
+ 0.013398447073996067,
1118
+ 0.038051582872867584
1119
+ ],
1120
+ "std": [
1121
+ 0.4350316822528839,
1122
+ 0.3735441565513611,
1123
+ 0.039987526834010384,
1124
+ 0.08149176836013794,
1125
+ 0.025776363909243462,
1126
+ 0.09133275598287582,
1127
+ 0.08737349510192871
1128
+ ],
1129
+ "q01": [
1130
+ 0.0,
1131
+ 0.0,
1132
+ 0.9262623804807663,
1133
+ -0.3768792775273323,
1134
+ 0.8234693193435669,
1135
+ -0.13829821586608887,
1136
+ 0.0
1137
+ ],
1138
+ "q99": [
1139
+ 1.0,
1140
+ 1.0,
1141
+ 1.0,
1142
+ 0.0434911641851068,
1143
+ 1.0,
1144
+ 0.5673608106374741,
1145
+ 0.3296095791459084
1146
+ ]
1147
+ }
1148
+ },
1149
+ "relative_action": {}
1150
+ }
1151
+ }
a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/final_model_config.json ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_type": "Gr00tN1d7",
3
+ "model_dtype": "bfloat16",
4
+ "model_name": "nvidia/Cosmos-Reason2-2B",
5
+ "backbone_model_type": "qwen",
6
+ "model_revision": null,
7
+ "tune_top_llm_layers": 0,
8
+ "backbone_embedding_dim": 2048,
9
+ "tune_llm": true,
10
+ "tune_visual": true,
11
+ "select_layer": 16,
12
+ "reproject_vision": false,
13
+ "use_flash_attention": true,
14
+ "load_bf16": false,
15
+ "backbone_trainable_params_fp32": true,
16
+ "letter_box_transform": false,
17
+ "extra_augmentation_config": null,
18
+ "apply_sincos_state_encoding": false,
19
+ "use_percentiles": true,
20
+ "use_relative_action": false,
21
+ "max_state_dim": 132,
22
+ "max_action_dim": 132,
23
+ "action_horizon": 40,
24
+ "strict_action_padding_mask": false,
25
+ "hidden_size": 1024,
26
+ "input_embedding_dim": 1536,
27
+ "state_history_length": 1,
28
+ "add_pos_embed": true,
29
+ "attn_dropout": 0.2,
30
+ "use_vlln": true,
31
+ "strict_vl_padding_mask": false,
32
+ "max_seq_len": 1024,
33
+ "use_alternate_vl_dit": true,
34
+ "attend_text_every_n_blocks": 2,
35
+ "diffusion_model_cfg": {
36
+ "attention_head_dim": 48,
37
+ "dropout": 0.2,
38
+ "final_dropout": true,
39
+ "interleave_self_attention": true,
40
+ "norm_type": "ada_norm",
41
+ "num_attention_heads": 32,
42
+ "num_layers": 32,
43
+ "output_dim": 1024,
44
+ "positional_embeddings": null
45
+ },
46
+ "num_inference_timesteps": 4,
47
+ "noise_beta_alpha": 1.5,
48
+ "noise_beta_beta": 1.0,
49
+ "noise_s": 0.999,
50
+ "num_timestep_buckets": 1000,
51
+ "tune_projector": true,
52
+ "tune_diffusion_model": true,
53
+ "tune_vlln": true,
54
+ "state_dropout_prob": 0.2,
55
+ "exclude_state": false,
56
+ "use_mean_std": false,
57
+ "max_num_embodiments": 32
58
+ }
a14-unscrew-bulb/v2-b512/step-002000/experiment_cfg/final_processor_config.json ADDED
The diff for this file is too large to render. See raw diff
 
a14-unscrew-bulb/v2-b512/step-002000/loss_curve.csv ADDED
@@ -0,0 +1,401 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ step,loss,learning_rate,grad_norm
2
+ 10,1.4311,4.5e-06,0.45548781752586365
3
+ 20,1.3387,9.5e-06,0.6781157851219177
4
+ 30,1.2232,1.45e-05,0.603681743144989
5
+ 40,1.1367,1.9500000000000003e-05,0.5604022741317749
6
+ 50,1.0905,2.45e-05,1.2952625751495361
7
+ 60,1.0696,2.95e-05,0.6026202440261841
8
+ 70,1.0596,3.45e-05,0.3096122145652771
9
+ 80,1.0537,3.9500000000000005e-05,0.2874182462692261
10
+ 90,1.0422,4.4500000000000004e-05,0.2818102240562439
11
+ 100,1.0423,4.9500000000000004e-05,0.4246894121170044
12
+ 110,1.0333,5.45e-05,0.5688769817352295
13
+ 120,0.9992,5.95e-05,0.5930460691452026
14
+ 130,0.9417,6.450000000000001e-05,0.5382891893386841
15
+ 140,0.8895,6.95e-05,0.7342086434364319
16
+ 150,0.8431,7.450000000000001e-05,0.8100906014442444
17
+ 160,0.8041,7.950000000000001e-05,0.6786221861839294
18
+ 170,0.7719,8.450000000000001e-05,0.8967251777648926
19
+ 180,0.7539,8.950000000000001e-05,0.7777813673019409
20
+ 190,0.7309,9.449999999999999e-05,0.8444279432296753
21
+ 200,0.718,9.95e-05,0.7900912165641785
22
+ 210,0.7104,9.999861593790126e-05,0.6578666567802429
23
+ 220,0.7047,9.999383162408304e-05,0.7282286882400513
24
+ 230,0.7024,9.998563029828259e-05,0.7857356667518616
25
+ 240,0.6952,9.997401252104962e-05,0.7012117505073547
26
+ 250,0.6898,9.995897908644378e-05,0.6767271161079407
27
+ 260,0.6809,9.994053102198034e-05,0.8207566142082214
28
+ 270,0.6699,9.991866958856003e-05,0.9090561866760254
29
+ 280,0.6438,9.989339628038276e-05,0.833543598651886
30
+ 290,0.6349,9.98647128248456e-05,0.7834540605545044
31
+ 300,0.6051,9.98326211824246e-05,0.9627916216850281
32
+ 310,0.5885,9.979712354654091e-05,0.7982354760169983
33
+ 320,0.5621,9.975822234341079e-05,0.7507064342498779
34
+ 330,0.5466,9.97159202318798e-05,1.0718227624893188
35
+ 340,0.5236,9.967022010324105e-05,1.0223731994628906
36
+ 350,0.5016,9.962112508103765e-05,0.8068194389343262
37
+ 360,0.484,9.956863852084914e-05,1.0485377311706543
38
+ 370,0.4736,9.951276401006221e-05,0.8841873407363892
39
+ 380,0.4506,9.945350536762543e-05,1.0380600690841675
40
+ 390,0.4317,9.939086664378829e-05,1.1374574899673462
41
+ 400,0.4166,9.932485211982437e-05,1.1841020584106445
42
+ 410,0.4057,9.92554663077387e-05,1.1365927457809448
43
+ 420,0.3918,9.918271394995935e-05,1.2993481159210205
44
+ 430,0.3842,9.910660001901335e-05,1.1606972217559814
45
+ 440,0.3623,9.902712971718675e-05,1.272127628326416
46
+ 450,0.3492,9.894430847616915e-05,1.0958130359649658
47
+ 460,0.3361,9.885814195668232e-05,1.047905683517456
48
+ 470,0.3306,9.876863604809344e-05,1.2776614427566528
49
+ 480,0.3091,9.867579686801245e-05,1.2377774715423584
50
+ 490,0.2942,9.8579630761874e-05,1.2725355625152588
51
+ 500,0.2801,9.848014430250367e-05,1.1680372953414917
52
+ 510,0.2755,9.837734428966885e-05,1.1904488801956177
53
+ 520,0.2507,9.827123774961383e-05,1.2455122470855713
54
+ 530,0.249,9.816183193457968e-05,1.3304517269134521
55
+ 540,0.2308,9.804913432230856e-05,1.105523943901062
56
+ 550,0.2225,9.793315261553252e-05,1.0879801511764526
57
+ 560,0.2076,9.781389474144717e-05,1.3669681549072266
58
+ 570,0.193,9.76913688511698e-05,1.2130101919174194
59
+ 580,0.1868,9.756558331918227e-05,1.3893616199493408
60
+ 590,0.1777,9.743654674275855e-05,1.3393754959106445
61
+ 600,0.1676,9.730426794137727e-05,1.2126128673553467
62
+ 610,0.1637,9.716875595611879e-05,1.4670941829681396
63
+ 620,0.1378,9.703002004904729e-05,1.126127004623413
64
+ 630,0.1278,9.688806970257773e-05,1.18602454662323
65
+ 640,0.1231,9.674291461882774e-05,1.3860094547271729
66
+ 650,0.1235,9.659456471895445e-05,1.3980443477630615
67
+ 660,0.1123,9.644303014247648e-05,1.209280252456665
68
+ 670,0.1269,9.628832124658085e-05,1.2374107837677002
69
+ 680,0.1101,9.613044860541507e-05,1.175325632095337
70
+ 690,0.1031,9.596942300936445e-05,1.1790177822113037
71
+ 700,0.1052,9.580525546431459e-05,1.1817500591278076
72
+ 710,0.1116,9.563795719089911e-05,1.258915901184082
73
+ 720,0.0937,9.546753962373281e-05,1.2639306783676147
74
+ 730,0.0928,9.529401441062997e-05,1.1046220064163208
75
+ 740,0.0788,9.511739341180842e-05,1.2190810441970825
76
+ 750,0.0894,9.493768869907886e-05,1.2571676969528198
77
+ 760,0.0807,9.475491255501968e-05,1.0401405096054077
78
+ 770,0.0829,9.456907747213748e-05,1.1861028671264648
79
+ 780,0.0843,9.438019615201336e-05,1.1651418209075928
80
+ 790,0.0788,9.418828150443469e-05,0.9502962827682495
81
+ 800,0.0842,9.399334664651262e-05,1.1027251482009888
82
+ 810,0.088,9.379540490178581e-05,1.1836992502212524
83
+ 820,0.0818,9.359446979930955e-05,1.1526118516921997
84
+ 830,0.0822,9.33905550727312e-05,1.2146070003509521
85
+ 840,0.0762,9.318367465935142e-05,1.1088842153549194
86
+ 850,0.0873,9.29738426991717e-05,1.0168155431747437
87
+ 860,0.0771,9.276107353392774e-05,1.0673962831497192
88
+ 870,0.0773,9.254538170610938e-05,1.0738420486450195
89
+ 880,0.0721,9.232678195796654e-05,1.0947306156158447
90
+ 890,0.0691,9.210528923050164e-05,1.12360417842865
91
+ 900,0.0812,9.188091866244834e-05,1.0733106136322021
92
+ 910,0.0654,9.165368558923695e-05,1.0893988609313965
93
+ 920,0.0753,9.142360554194618e-05,1.1471936702728271
94
+ 930,0.0682,9.119069424624163e-05,1.1242029666900635
95
+ 940,0.0655,9.0954967621301e-05,1.0201963186264038
96
+ 950,0.0763,9.071644177872594e-05,1.1408215761184692
97
+ 960,0.0622,9.047513302144095e-05,1.054488182067871
98
+ 970,0.0644,9.023105784257906e-05,1.0666213035583496
99
+ 980,0.0695,8.998423292435454e-05,1.07891047000885
100
+ 990,0.0704,8.973467513692265e-05,1.0655622482299805
101
+ 1000,0.0674,8.94824015372267e-05,1.116550326347351
102
+ 1010,0.0631,8.922742936783207e-05,0.8838066458702087
103
+ 1020,0.0665,8.896977605574788e-05,1.086512565612793
104
+ 1030,0.0653,8.870945921123576e-05,1.0114772319793701
105
+ 1040,0.067,8.844649662660624e-05,1.0343341827392578
106
+ 1050,0.0594,8.818090627500266e-05,0.9564461708068848
107
+ 1060,0.0734,8.791270630917275e-05,0.8460029363632202
108
+ 1070,0.0675,8.764191506022795e-05,1.006821870803833
109
+ 1080,0.0678,8.736855103639037e-05,0.9695994257926941
110
+ 1090,0.0532,8.709263292172794e-05,1.0134460926055908
111
+ 1100,0.0612,8.681417957487729e-05,1.0870625972747803
112
+ 1110,0.0664,8.653321002775478e-05,1.0378804206848145
113
+ 1120,0.0661,8.624974348425574e-05,1.0451186895370483
114
+ 1130,0.0711,8.596379931894188e-05,1.0146639347076416
115
+ 1140,0.0682,8.567539707571703e-05,0.9299578070640564
116
+ 1150,0.0652,8.538455646649146e-05,1.147046446800232
117
+ 1160,0.0623,8.509129736983446e-05,1.0006110668182373
118
+ 1170,0.0587,8.479563982961571e-05,0.9463521838188171
119
+ 1180,0.0605,8.449760405363539e-05,1.073960781097412
120
+ 1190,0.0629,8.419721041224287e-05,1.012939214706421
121
+ 1200,0.065,8.389447943694451e-05,0.8890576958656311
122
+ 1210,0.0614,8.358943181900032e-05,0.9021549820899963
123
+ 1220,0.0556,8.328208840800981e-05,0.9169679880142212
124
+ 1230,0.0624,8.297247021048686e-05,1.0166311264038086
125
+ 1240,0.0556,8.266059838842396e-05,0.9880717396736145
126
+ 1250,0.062,8.23464942578459e-05,1.0422134399414062
127
+ 1260,0.0549,8.203017928735277e-05,0.9401381611824036
128
+ 1270,0.0602,8.17116750966526e-05,1.0414115190505981
129
+ 1280,0.0582,8.139100345508377e-05,0.9322880506515503
130
+ 1290,0.0602,8.106818628012697e-05,1.06840980052948
131
+ 1300,0.0564,8.074324563590736e-05,1.0114227533340454
132
+ 1310,0.0534,8.041620373168628e-05,0.93513423204422
133
+ 1320,0.0585,8.008708292034349e-05,0.9780723452568054
134
+ 1330,0.0513,7.975590569684925e-05,0.8210321068763733
135
+ 1340,0.0541,7.942269469672687e-05,0.9770547151565552
136
+ 1350,0.0487,7.908747269450558e-05,1.064526081085205
137
+ 1360,0.0527,7.875026260216393e-05,0.8699714541435242
138
+ 1370,0.0439,7.841108746756382e-05,0.8412991762161255
139
+ 1380,0.0492,7.806997047287516e-05,0.8587125539779663
140
+ 1390,0.0457,7.772693493299138e-05,1.053528070449829
141
+ 1400,0.0549,7.7382004293936e-05,0.9603511691093445
142
+ 1410,0.0456,7.703520213126e-05,0.9119125008583069
143
+ 1420,0.0537,7.66865521484305e-05,0.8282575607299805
144
+ 1430,0.0504,7.633607817521074e-05,0.8297218084335327
145
+ 1440,0.051,7.598380416603119e-05,0.9777206778526306
146
+ 1450,0.0536,7.562975419835247e-05,0.8542219400405884
147
+ 1460,0.0504,7.527395247101956e-05,0.8440294861793518
148
+ 1470,0.0502,7.491642330260789e-05,0.7946062684059143
149
+ 1480,0.0563,7.45571911297612e-05,0.9157381057739258
150
+ 1490,0.046,7.419628050552131e-05,0.8264444470405579
151
+ 1500,0.0591,7.383371609764999e-05,0.9470329880714417
152
+ 1510,0.0482,7.346952268694288e-05,0.7743332982063293
153
+ 1520,0.0495,7.310372516553585e-05,1.101076364517212
154
+ 1530,0.0497,7.273634853520356e-05,0.8662927746772766
155
+ 1540,0.0563,7.236741790565072e-05,0.8429809808731079
156
+ 1550,0.052,7.199695849279576e-05,0.9624887704849243
157
+ 1560,0.048,7.162499561704747e-05,0.887871265411377
158
+ 1570,0.0459,7.125155470157429e-05,0.9352372288703918
159
+ 1580,0.044,7.087666127056675e-05,0.8238629698753357
160
+ 1590,0.0476,7.050034094749286e-05,0.8751165866851807
161
+ 1600,0.0461,7.012261945334683e-05,0.8804195523262024
162
+ 1610,0.0402,6.974352260489103e-05,0.7688097357749939
163
+ 1620,0.0401,6.936307631289148e-05,0.6759632229804993
164
+ 1630,0.0407,6.898130658034685e-05,0.8520809412002563
165
+ 1640,0.046,6.859823950071127e-05,1.0360755920410156
166
+ 1650,0.0504,6.821390125611078e-05,0.9505870938301086
167
+ 1660,0.037,6.782831811555385e-05,0.7323496341705322
168
+ 1670,0.0441,6.744151643313597e-05,0.8076654076576233
169
+ 1680,0.048,6.705352264623828e-05,0.8517192602157593
170
+ 1690,0.048,6.666436327372078e-05,0.8317267894744873
171
+ 1700,0.0439,6.62740649141096e-05,0.9000750184059143
172
+ 1710,0.043,6.588265424377919e-05,0.8319242000579834
173
+ 1720,0.0473,6.549015801512895e-05,0.6666507720947266
174
+ 1730,0.045,6.509660305475468e-05,0.8333040475845337
175
+ 1740,0.0396,6.47020162616152e-05,0.8183309435844421
176
+ 1750,0.0428,6.430642460519365e-05,0.9074743390083313
177
+ 1760,0.0477,6.390985512365426e-05,0.8012915253639221
178
+ 1770,0.044,6.351233492199431e-05,0.7697868943214417
179
+ 1780,0.0422,6.311389117019155e-05,0.7668655514717102
180
+ 1790,0.0476,6.271455110134713e-05,0.8269855380058289
181
+ 1800,0.0429,6.231434200982428e-05,0.8340345025062561
182
+ 1810,0.0411,6.191329124938285e-05,0.8056420683860779
183
+ 1820,0.0375,6.15114262313095e-05,0.7833528518676758
184
+ 1830,0.0379,6.110877442254444e-05,0.7766333818435669
185
+ 1840,0.0379,6.0705363343803946e-05,0.8195701241493225
186
+ 1850,0.0387,6.030122056769934e-05,0.7879520654678345
187
+ 1860,0.0443,5.989637371685257e-05,0.8573419451713562
188
+ 1870,0.0442,5.949085046200808e-05,0.6557536721229553
189
+ 1880,0.0362,5.908467852014169e-05,0.7351909279823303
190
+ 1890,0.0364,5.867788565256607e-05,0.7618509531021118
191
+ 1900,0.0482,5.827049966303335e-05,0.8200004696846008
192
+ 1910,0.0421,5.786254839583478e-05,0.847327709197998
193
+ 1920,0.0394,5.745405973389757e-05,0.8439093828201294
194
+ 1930,0.0411,5.7045061596879134e-05,0.7736167907714844
195
+ 1940,0.0397,5.6635581939258855e-05,0.8771815896034241
196
+ 1950,0.0404,5.622564874842742e-05,0.7791712284088135
197
+ 1960,0.0337,5.5815290042773836e-05,0.8065997362136841
198
+ 1970,0.041,5.540453386977058e-05,0.8823350667953491
199
+ 1980,0.0488,5.4993408304056425e-05,0.7548691630363464
200
+ 1990,0.0377,5.458194144551768e-05,0.7795559167861938
201
+ 2000,0.034,5.417016141736756e-05,0.7108629941940308
202
+ 2010,0.0359,5.375809636422399e-05,0.7524849772453308
203
+ 2020,0.0383,5.334577445018599e-05,0.7699241042137146
204
+ 2030,0.041,5.293322385690867e-05,0.76553875207901
205
+ 2040,0.0434,5.252047278167709e-05,0.7431535720825195
206
+ 2050,0.0367,5.210754943547893e-05,0.6975953578948975
207
+ 2060,0.0341,5.169448204107643e-05,0.6527072787284851
208
+ 2070,0.0341,5.128129883107729e-05,0.6326196789741516
209
+ 2080,0.0323,5.086802804600505e-05,0.5768119096755981
210
+ 2090,0.0337,5.045469793236892e-05,0.7889487147331238
211
+ 2100,0.0325,5.00413367407331e-05,0.7711167335510254
212
+ 2110,0.0334,4.9627972723785964e-05,0.8419650793075562
213
+ 2120,0.0314,4.921463413440898e-05,0.7012775540351868
214
+ 2130,0.0338,4.8801349223745654e-05,0.5781388878822327
215
+ 2140,0.0344,4.838814623927067e-05,0.6582086086273193
216
+ 2150,0.042,4.797505342285912e-05,0.6766339540481567
217
+ 2160,0.0339,4.756209900885628e-05,0.6903968453407288
218
+ 2170,0.0314,4.714931122214781e-05,0.543709397315979
219
+ 2180,0.0366,4.673671827623058e-05,0.7570103406906128
220
+ 2190,0.0316,4.632434837128443e-05,0.6662298440933228
221
+ 2200,0.0318,4.591222969224453e-05,0.550093948841095
222
+ 2210,0.0293,4.550039040687518e-05,0.603888213634491
223
+ 2220,0.0313,4.508885866384446e-05,0.662289559841156
224
+ 2230,0.0349,4.4677662590800355e-05,0.7093011736869812
225
+ 2240,0.0304,4.426683029244825e-05,0.6487430930137634
226
+ 2250,0.0358,4.385638984863e-05,0.6345328092575073
227
+ 2260,0.0352,4.3446369312404745e-05,0.7010223865509033
228
+ 2270,0.0302,4.3036796708131474e-05,0.6635097861289978
229
+ 2280,0.0363,4.262770002955363e-05,0.726793646812439
230
+ 2290,0.0327,4.221910723788578e-05,0.6263322234153748
231
+ 2300,0.0286,4.1811046259902474e-05,0.6477052569389343
232
+ 2310,0.0388,4.140354498602952e-05,0.6114294528961182
233
+ 2320,0.0302,4.099663126843769e-05,0.6365948915481567
234
+ 2330,0.0392,4.059033291913902e-05,0.6756660342216492
235
+ 2340,0.0305,4.0184677708086014e-05,0.6872760057449341
236
+ 2350,0.0354,3.977969336127348e-05,0.5725398659706116
237
+ 2360,0.0302,3.937540755884357e-05,0.6011505126953125
238
+ 2370,0.031,3.897184793319384e-05,0.6561709046363831
239
+ 2380,0.0287,3.856904206708863e-05,0.57468181848526
240
+ 2390,0.0352,3.8167017491773847e-05,0.6154991388320923
241
+ 2400,0.0351,3.776580168509516e-05,0.7515709400177002
242
+ 2410,0.029,3.736542206962e-05,0.5898580551147461
243
+ 2420,0.0307,3.696590601076326e-05,0.6231646537780762
244
+ 2430,0.0274,3.656728081491686e-05,0.566853404045105
245
+ 2440,0.0326,3.6169573727583405e-05,0.6385308504104614
246
+ 2450,0.0289,3.5772811931514036e-05,0.6533120274543762
247
+ 2460,0.035,3.5377022544850505e-05,0.7482478618621826
248
+ 2470,0.0307,3.498223261927158e-05,0.6108620762825012
249
+ 2480,0.0304,3.4588469138144295e-05,0.7000325918197632
250
+ 2490,0.0254,3.419575901467952e-05,0.5968828797340393
251
+ 2500,0.033,3.380412909009254e-05,0.6828427910804749
252
+ 2510,0.0309,3.3413606131768475e-05,0.680059552192688
253
+ 2520,0.0294,3.302421683143279e-05,0.5332454442977905
254
+ 2530,0.0258,3.2635987803326896e-05,0.5806804895401001
255
+ 2540,0.0318,3.224894558238918e-05,0.6283760666847229
256
+ 2550,0.0368,3.18631166224413e-05,0.7224867343902588
257
+ 2560,0.0312,3.147852729438017e-05,0.6222689747810364
258
+ 2570,0.0284,3.109520388437548e-05,0.607688307762146
259
+ 2580,0.0307,3.0713172592073116e-05,0.6588331460952759
260
+ 2590,0.03,3.0332459528804457e-05,0.6146456599235535
261
+ 2600,0.0272,2.9953090715801634e-05,0.6048339009284973
262
+ 2610,0.0264,2.9575092082419086e-05,0.5202537178993225
263
+ 2620,0.0284,2.9198489464361288e-05,0.5415447354316711
264
+ 2630,0.0258,2.8823308601916948e-05,0.4601195454597473
265
+ 2640,0.0239,2.8449575138199613e-05,0.6900708675384521
266
+ 2650,0.0301,2.807731461739509e-05,0.5642911791801453
267
+ 2660,0.0378,2.7706552483015485e-05,0.682118833065033
268
+ 2670,0.0313,2.733731407616018e-05,0.47676756978034973
269
+ 2680,0.0336,2.6969624633783806e-05,0.505494236946106
270
+ 2690,0.0268,2.660350928697134e-05,0.5062770247459412
271
+ 2700,0.0319,2.6238993059220395e-05,0.6475841999053955
272
+ 2710,0.0298,2.5876100864730933e-05,0.5577534437179565
273
+ 2720,0.0264,2.5514857506702405e-05,0.5197186470031738
274
+ 2730,0.0276,2.5155287675638474e-05,0.664161205291748
275
+ 2740,0.0294,2.4797415947659457e-05,0.5872769355773926
276
+ 2750,0.0241,2.4441266782822588e-05,0.46903419494628906
277
+ 2760,0.026,2.4086864523450183e-05,0.5618360042572021
278
+ 2770,0.027,2.3734233392465903e-05,0.4858380854129791
279
+ 2780,0.0271,2.3383397491739145e-05,0.546043872833252
280
+ 2790,0.0247,2.3034380800437678e-05,0.4840218126773834
281
+ 2800,0.0263,2.2687207173388743e-05,0.5970568060874939
282
+ 2810,0.0287,2.234190033944858e-05,0.5109860301017761
283
+ 2820,0.0236,2.1998483899880596e-05,0.491593599319458
284
+ 2830,0.0264,2.1656981326742266e-05,0.4333474040031433
285
+ 2840,0.0272,2.1317415961280824e-05,0.4792613983154297
286
+ 2850,0.0268,2.097981101233794e-05,0.6060649752616882
287
+ 2860,0.027,2.0644189554763417e-05,0.5738887190818787
288
+ 2870,0.0257,2.0310574527838072e-05,0.5344540476799011
289
+ 2880,0.0241,1.9978988733705807e-05,0.5391669869422913
290
+ 2890,0.0238,1.9649454835815202e-05,0.41440337896347046
291
+ 2900,0.0254,1.932199535737045e-05,0.5010972023010254
292
+ 2910,0.0371,1.8996632679791914e-05,0.6001975536346436
293
+ 2920,0.0244,1.8673389041186418e-05,0.6123016476631165
294
+ 2930,0.0252,1.8352286534827274e-05,0.512846827507019
295
+ 2940,0.0251,1.803334710764426e-05,0.39674296975135803
296
+ 2950,0.0247,1.7716592558723556e-05,0.41568347811698914
297
+ 2960,0.03,1.7402044537817824e-05,0.5356529951095581
298
+ 2970,0.0286,1.7089724543866465e-05,0.4542825222015381
299
+ 2980,0.0233,1.6779653923526188e-05,0.387778103351593
300
+ 2990,0.0217,1.6471853869712023e-05,0.45108118653297424
301
+ 3000,0.0281,1.6166345420148787e-05,0.4745352864265442
302
+ 3010,0.0289,1.5863149455933158e-05,0.550599217414856
303
+ 3020,0.0311,1.5562286700106558e-05,0.5312296748161316
304
+ 3030,0.0231,1.526377771623869e-05,0.4456547498703003
305
+ 3040,0.0222,1.496764290702209e-05,0.38005807995796204
306
+ 3050,0.0272,1.4673902512877585e-05,0.4038347899913788
307
+ 3060,0.0268,1.438257661057093e-05,0.46522170305252075
308
+ 3070,0.029,1.4093685111840566e-05,0.39163708686828613
309
+ 3080,0.0248,1.380724776203668e-05,0.4121502935886383
310
+ 3090,0.0249,1.3523284138771642e-05,0.3686705231666565
311
+ 3100,0.0248,1.3241813650581902e-05,0.39747363328933716
312
+ 3110,0.0293,1.2962855535601438e-05,0.49244391918182373
313
+ 3120,0.0232,1.2686428860246852e-05,0.3457881808280945
314
+ 3130,0.0265,1.241255251791421e-05,0.518446147441864
315
+ 3140,0.0252,1.2141245227687731e-05,0.35663512349128723
316
+ 3150,0.0249,1.1872525533060269e-05,0.42742785811424255
317
+ 3160,0.0221,1.1606411800666028e-05,0.43492045998573303
318
+ 3170,0.0267,1.134292221902511e-05,0.4053647518157959
319
+ 3180,0.0219,1.1082074797300413e-05,0.419670432806015
320
+ 3190,0.0252,1.0823887364066737e-05,0.42409881949424744
321
+ 3200,0.0281,1.0568377566092164e-05,0.5156289339065552
322
+ 3210,0.0255,1.0315562867131983e-05,0.4180368483066559
323
+ 3220,0.0218,1.0065460546735045e-05,0.39815619587898254
324
+ 3230,0.023,9.818087699062716e-06,0.3304533362388611
325
+ 3240,0.0197,9.57346123172055e-06,0.2590312957763672
326
+ 3250,0.0197,9.331597864602632e-06,0.30036982893943787
327
+ 3260,0.0314,9.09251412874882e-06,0.34973791241645813
328
+ 3270,0.0239,8.856226365214897e-06,0.3463924825191498
329
+ 3280,0.0215,8.622750723955597e-06,0.3705339729785919
330
+ 3290,0.0231,8.392103162720883e-06,0.3311239778995514
331
+ 3300,0.0212,8.164299445965167e-06,0.40537145733833313
332
+ 3310,0.022,7.939355143769905e-06,0.3530327081680298
333
+ 3320,0.0199,7.717285630779341e-06,0.311110258102417
334
+ 3330,0.0228,7.498106085149697e-06,0.4062057137489319
335
+ 3340,0.0223,7.2818314875117755e-06,0.36281439661979675
336
+ 3350,0.0282,7.06847661994704e-06,0.4808196723461151
337
+ 3360,0.0211,6.85805606497727e-06,0.32958048582077026
338
+ 3370,0.0323,6.650584204567889e-06,0.35616374015808105
339
+ 3380,0.0262,6.446075219144965e-06,0.39464375376701355
340
+ 3390,0.0259,6.244543086625987e-06,0.3493010699748993
341
+ 3400,0.0285,6.046001581464505e-06,0.36447426676750183
342
+ 3410,0.0237,5.850464273708689e-06,0.36249032616615295
343
+ 3420,0.0232,5.657944528073745e-06,0.31157174706459045
344
+ 3430,0.0212,5.468455503028574e-06,0.2522493302822113
345
+ 3440,0.0245,5.282010149896327e-06,0.38474592566490173
346
+ 3450,0.02,5.098621211969223e-06,0.2883252501487732
347
+ 3460,0.0259,4.918301223637573e-06,0.33924245834350586
348
+ 3470,0.0203,4.74106250953304e-06,0.22463372349739075
349
+ 3480,0.0235,4.566917183686309e-06,0.28400376439094543
350
+ 3490,0.0222,4.3958771486990735e-06,0.2881617844104767
351
+ 3500,0.0296,4.22795409493052e-06,0.3541630804538727
352
+ 3510,0.0238,4.06315949969831e-06,0.22167044878005981
353
+ 3520,0.0218,3.901504626494135e-06,0.24586635828018188
354
+ 3530,0.0224,3.7430005242138354e-06,0.30878713726997375
355
+ 3540,0.0257,3.587658026402263e-06,0.21205803751945496
356
+ 3550,0.0241,3.435487750512778e-06,0.25614380836486816
357
+ 3560,0.0231,3.2865000971816107e-06,0.2831701338291168
358
+ 3570,0.0249,3.1407052495169566e-06,0.3652791380882263
359
+ 3580,0.0241,2.9981131724029887e-06,0.20656822621822357
360
+ 3590,0.0213,2.858733611818765e-06,0.19780467450618744
361
+ 3600,0.0245,2.7225760941721136e-06,0.2644926905632019
362
+ 3610,0.0194,2.589649925648491e-06,0.2240334153175354
363
+ 3620,0.022,2.459964191574948e-06,0.20334111154079437
364
+ 3630,0.0225,2.3335277557991364e-06,0.2289544939994812
365
+ 3640,0.0242,2.210349260083494e-06,0.2667798101902008
366
+ 3650,0.026,2.0904371235145827e-06,0.20558719336986542
367
+ 3660,0.0186,1.9737995419276455e-06,0.20043709874153137
368
+ 3670,0.0213,1.8604444873464466e-06,0.2401365041732788
369
+ 3680,0.0253,1.7503797074383988e-06,0.2516518533229828
370
+ 3690,0.025,1.6436127249849998e-06,0.26054397225379944
371
+ 3700,0.0202,1.5401508373676764e-06,0.26021912693977356
372
+ 3710,0.0275,1.4400011160690175e-06,0.1831081211566925
373
+ 3720,0.0241,1.3431704061894312e-06,0.19615495204925537
374
+ 3730,0.0204,1.2496653259793268e-06,0.22023119032382965
375
+ 3740,0.0269,1.1594922663867135e-06,0.23225434124469757
376
+ 3750,0.023,1.0726573906204463e-06,0.24306732416152954
377
+ 3760,0.0222,9.891666337289273e-07,0.12998664379119873
378
+ 3770,0.0277,9.090257021944881e-07,0.2530870735645294
379
+ 3780,0.0215,8.3224007354335e-07,0.13551126420497894
380
+ 3790,0.0205,7.588149959712243e-07,0.224202960729599
381
+ 3800,0.0241,6.887554879846326e-07,0.20291346311569214
382
+ 3810,0.0195,6.220663380578861e-07,0.20860302448272705
383
+ 3820,0.0244,5.58752104305793e-07,0.11800605058670044
384
+ 3830,0.0251,4.988171141721232e-07,0.13866862654685974
385
+ 3840,0.0215,4.4226546413383974e-07,0.1725994348526001
386
+ 3850,0.0214,3.891010194211009e-07,0.18068577349185944
387
+ 3860,0.0243,3.393274137530877e-07,0.14011624455451965
388
+ 3870,0.0232,2.9294804908962525e-07,0.2885558307170868
389
+ 3880,0.0226,2.499660953986849e-07,0.1305045336484909
390
+ 3890,0.0199,2.103844904397023e-07,0.1302751898765564
391
+ 3900,0.0233,1.742059395627993e-07,0.14684940874576569
392
+ 3910,0.0214,1.414329155238703e-07,0.18236541748046875
393
+ 3920,0.0209,1.1206765831557886e-07,0.14119434356689453
394
+ 3930,0.0216,8.611217501423574e-08,0.17236880958080292
395
+ 3940,0.0242,6.356823964266401e-08,0.1929444521665573
396
+ 3950,0.0253,4.4437393048885056e-08,0.20120930671691895
397
+ 3960,0.0199,2.8720942800858353e-08,0.1829187124967575
398
+ 3970,0.0245,1.6419963097080715e-08,0.11942435801029205
399
+ 3980,0.0206,7.535294693172822e-09,0.26055794954299927
400
+ 3990,0.0206,2.0675448444251732e-09,0.10503558069467545
401
+ 4000,0.0234,1.7087264264636914e-11,0.17020879685878754
a14-unscrew-bulb/v2-b512/step-002000/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:48a50785e27a0a41151ebcc97a558c514ba002aceb90aca84d745a857faa8c21
3
+ size 4990519232
a14-unscrew-bulb/v2-b512/step-002000/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1ddd720b3d4c2d9e6f3dd6e3d776c87f360bf84be86f2f60c0aca34d472a11cc
3
+ size 1919980184
a14-unscrew-bulb/v2-b512/step-002000/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
a14-unscrew-bulb/v2-b512/step-002000/processor_config.json ADDED
@@ -0,0 +1,1145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "processor_class": "Gr00tN1d7Processor",
3
+ "processor_kwargs": {
4
+ "modality_configs": {
5
+ "real_g1_relative_eef_relative_joints": {
6
+ "video": {
7
+ "delta_indices": [
8
+ -20,
9
+ 0
10
+ ],
11
+ "modality_keys": [
12
+ "ego_view"
13
+ ],
14
+ "sin_cos_embedding_keys": null,
15
+ "mean_std_embedding_keys": null,
16
+ "action_configs": null
17
+ },
18
+ "state": {
19
+ "delta_indices": [
20
+ 0
21
+ ],
22
+ "modality_keys": [
23
+ "left_wrist_eef_9d",
24
+ "right_wrist_eef_9d",
25
+ "left_hand",
26
+ "right_hand",
27
+ "left_arm",
28
+ "right_arm",
29
+ "waist"
30
+ ],
31
+ "sin_cos_embedding_keys": null,
32
+ "mean_std_embedding_keys": null,
33
+ "action_configs": null
34
+ },
35
+ "action": {
36
+ "delta_indices": [
37
+ 0,
38
+ 1,
39
+ 2,
40
+ 3,
41
+ 4,
42
+ 5,
43
+ 6,
44
+ 7,
45
+ 8,
46
+ 9,
47
+ 10,
48
+ 11,
49
+ 12,
50
+ 13,
51
+ 14,
52
+ 15,
53
+ 16,
54
+ 17,
55
+ 18,
56
+ 19,
57
+ 20,
58
+ 21,
59
+ 22,
60
+ 23,
61
+ 24,
62
+ 25,
63
+ 26,
64
+ 27,
65
+ 28,
66
+ 29,
67
+ 30,
68
+ 31,
69
+ 32,
70
+ 33,
71
+ 34,
72
+ 35,
73
+ 36,
74
+ 37,
75
+ 38,
76
+ 39
77
+ ],
78
+ "modality_keys": [
79
+ "left_wrist_eef_9d",
80
+ "right_wrist_eef_9d",
81
+ "left_hand",
82
+ "right_hand",
83
+ "left_arm",
84
+ "right_arm",
85
+ "waist",
86
+ "base_height_command",
87
+ "navigate_command"
88
+ ],
89
+ "sin_cos_embedding_keys": null,
90
+ "mean_std_embedding_keys": null,
91
+ "action_configs": [
92
+ {
93
+ "rep": "RELATIVE",
94
+ "type": "EEF",
95
+ "format": "XYZ_ROT6D",
96
+ "state_key": "left_wrist_eef_9d"
97
+ },
98
+ {
99
+ "rep": "RELATIVE",
100
+ "type": "EEF",
101
+ "format": "XYZ_ROT6D",
102
+ "state_key": "right_wrist_eef_9d"
103
+ },
104
+ {
105
+ "rep": "ABSOLUTE",
106
+ "type": "NON_EEF",
107
+ "format": "DEFAULT",
108
+ "state_key": "left_hand"
109
+ },
110
+ {
111
+ "rep": "ABSOLUTE",
112
+ "type": "NON_EEF",
113
+ "format": "DEFAULT",
114
+ "state_key": "right_hand"
115
+ },
116
+ {
117
+ "rep": "RELATIVE",
118
+ "type": "NON_EEF",
119
+ "format": "DEFAULT",
120
+ "state_key": "left_arm"
121
+ },
122
+ {
123
+ "rep": "RELATIVE",
124
+ "type": "NON_EEF",
125
+ "format": "DEFAULT",
126
+ "state_key": "right_arm"
127
+ },
128
+ {
129
+ "rep": "ABSOLUTE",
130
+ "type": "NON_EEF",
131
+ "format": "DEFAULT",
132
+ "state_key": "waist"
133
+ },
134
+ {
135
+ "rep": "ABSOLUTE",
136
+ "type": "NON_EEF",
137
+ "format": "DEFAULT",
138
+ "state_key": "base_height_command"
139
+ },
140
+ {
141
+ "rep": "ABSOLUTE",
142
+ "type": "NON_EEF",
143
+ "format": "DEFAULT",
144
+ "state_key": "navigate_command"
145
+ }
146
+ ]
147
+ },
148
+ "language": {
149
+ "delta_indices": [
150
+ 0
151
+ ],
152
+ "modality_keys": [
153
+ "annotation.human.task_description"
154
+ ],
155
+ "sin_cos_embedding_keys": null,
156
+ "mean_std_embedding_keys": null,
157
+ "action_configs": null
158
+ }
159
+ },
160
+ "real_r1_pro_sharpa_relative_eef_mecka": {
161
+ "video": {
162
+ "delta_indices": [
163
+ -30,
164
+ 0
165
+ ],
166
+ "modality_keys": [
167
+ "ego_view_cropratio_res320x240_freq30"
168
+ ],
169
+ "sin_cos_embedding_keys": null,
170
+ "mean_std_embedding_keys": null,
171
+ "action_configs": null
172
+ },
173
+ "state": {
174
+ "delta_indices": [
175
+ 0
176
+ ],
177
+ "modality_keys": [
178
+ "left_wrist_eef",
179
+ "right_wrist_eef",
180
+ "left_hand_joints",
181
+ "right_hand_joints"
182
+ ],
183
+ "sin_cos_embedding_keys": null,
184
+ "mean_std_embedding_keys": null,
185
+ "action_configs": null
186
+ },
187
+ "action": {
188
+ "delta_indices": [
189
+ 0,
190
+ 1,
191
+ 2,
192
+ 3,
193
+ 4,
194
+ 5,
195
+ 6,
196
+ 7,
197
+ 8,
198
+ 9,
199
+ 10,
200
+ 11,
201
+ 12,
202
+ 13,
203
+ 14,
204
+ 15,
205
+ 16,
206
+ 17,
207
+ 18,
208
+ 19,
209
+ 20,
210
+ 21,
211
+ 22,
212
+ 23,
213
+ 24,
214
+ 25,
215
+ 26,
216
+ 27,
217
+ 28,
218
+ 29,
219
+ 30,
220
+ 31,
221
+ 32,
222
+ 33,
223
+ 34,
224
+ 35,
225
+ 36,
226
+ 37,
227
+ 38,
228
+ 39
229
+ ],
230
+ "modality_keys": [
231
+ "left_wrist_eef",
232
+ "right_wrist_eef",
233
+ "left_hand_joints",
234
+ "right_hand_joints"
235
+ ],
236
+ "sin_cos_embedding_keys": null,
237
+ "mean_std_embedding_keys": null,
238
+ "action_configs": [
239
+ {
240
+ "rep": "RELATIVE",
241
+ "type": "EEF",
242
+ "format": "XYZ_ROT6D",
243
+ "state_key": "left_wrist_eef"
244
+ },
245
+ {
246
+ "rep": "RELATIVE",
247
+ "type": "EEF",
248
+ "format": "XYZ_ROT6D",
249
+ "state_key": "right_wrist_eef"
250
+ },
251
+ {
252
+ "rep": "ABSOLUTE",
253
+ "type": "NON_EEF",
254
+ "format": "DEFAULT",
255
+ "state_key": "left_hand_joints"
256
+ },
257
+ {
258
+ "rep": "ABSOLUTE",
259
+ "type": "NON_EEF",
260
+ "format": "DEFAULT",
261
+ "state_key": "right_hand_joints"
262
+ }
263
+ ]
264
+ },
265
+ "language": {
266
+ "delta_indices": [
267
+ 0
268
+ ],
269
+ "modality_keys": [
270
+ "annotation.human.coarse_action"
271
+ ],
272
+ "sin_cos_embedding_keys": null,
273
+ "mean_std_embedding_keys": null,
274
+ "action_configs": null
275
+ }
276
+ },
277
+ "real_r1_pro_sharpa_relative_eef_human": {
278
+ "video": {
279
+ "delta_indices": [
280
+ -20,
281
+ 0
282
+ ],
283
+ "modality_keys": [
284
+ "ego_view_res320x240_freq20",
285
+ "left_wrist_view_res320x240_freq20",
286
+ "right_wrist_view_res320x240_freq20"
287
+ ],
288
+ "sin_cos_embedding_keys": null,
289
+ "mean_std_embedding_keys": null,
290
+ "action_configs": null
291
+ },
292
+ "state": {
293
+ "delta_indices": [
294
+ 0
295
+ ],
296
+ "modality_keys": [
297
+ "left_wrist_eef",
298
+ "right_wrist_eef",
299
+ "left_hand_joints",
300
+ "right_hand_joints"
301
+ ],
302
+ "sin_cos_embedding_keys": null,
303
+ "mean_std_embedding_keys": null,
304
+ "action_configs": null
305
+ },
306
+ "action": {
307
+ "delta_indices": [
308
+ 0,
309
+ 1,
310
+ 2,
311
+ 3,
312
+ 4,
313
+ 5,
314
+ 6,
315
+ 7,
316
+ 8,
317
+ 9,
318
+ 10,
319
+ 11,
320
+ 12,
321
+ 13,
322
+ 14,
323
+ 15,
324
+ 16,
325
+ 17,
326
+ 18,
327
+ 19,
328
+ 20,
329
+ 21,
330
+ 22,
331
+ 23,
332
+ 24,
333
+ 25,
334
+ 26,
335
+ 27,
336
+ 28,
337
+ 29,
338
+ 30,
339
+ 31,
340
+ 32,
341
+ 33,
342
+ 34,
343
+ 35,
344
+ 36,
345
+ 37,
346
+ 38,
347
+ 39
348
+ ],
349
+ "modality_keys": [
350
+ "left_wrist_eef",
351
+ "right_wrist_eef",
352
+ "left_hand_joints",
353
+ "right_hand_joints"
354
+ ],
355
+ "sin_cos_embedding_keys": null,
356
+ "mean_std_embedding_keys": null,
357
+ "action_configs": [
358
+ {
359
+ "rep": "RELATIVE",
360
+ "type": "EEF",
361
+ "format": "XYZ_ROT6D",
362
+ "state_key": "left_wrist_eef"
363
+ },
364
+ {
365
+ "rep": "RELATIVE",
366
+ "type": "EEF",
367
+ "format": "XYZ_ROT6D",
368
+ "state_key": "right_wrist_eef"
369
+ },
370
+ {
371
+ "rep": "ABSOLUTE",
372
+ "type": "NON_EEF",
373
+ "format": "DEFAULT",
374
+ "state_key": "left_hand_joints"
375
+ },
376
+ {
377
+ "rep": "ABSOLUTE",
378
+ "type": "NON_EEF",
379
+ "format": "DEFAULT",
380
+ "state_key": "right_hand_joints"
381
+ }
382
+ ]
383
+ },
384
+ "language": {
385
+ "delta_indices": [
386
+ 0
387
+ ],
388
+ "modality_keys": [
389
+ "annotation.human.coarse_action"
390
+ ],
391
+ "sin_cos_embedding_keys": null,
392
+ "mean_std_embedding_keys": null,
393
+ "action_configs": null
394
+ }
395
+ },
396
+ "real_r1_pro_sharpa_relative_eef": {
397
+ "video": {
398
+ "delta_indices": [
399
+ -20,
400
+ 0
401
+ ],
402
+ "modality_keys": [
403
+ "ego_view_res320x240_freq20",
404
+ "left_wrist_view_res320x240_freq20",
405
+ "right_wrist_view_res320x240_freq20"
406
+ ],
407
+ "sin_cos_embedding_keys": null,
408
+ "mean_std_embedding_keys": null,
409
+ "action_configs": null
410
+ },
411
+ "state": {
412
+ "delta_indices": [
413
+ 0
414
+ ],
415
+ "modality_keys": [
416
+ "left_wrist_eef",
417
+ "right_wrist_eef",
418
+ "left_hand_joints",
419
+ "right_hand_joints"
420
+ ],
421
+ "sin_cos_embedding_keys": null,
422
+ "mean_std_embedding_keys": null,
423
+ "action_configs": null
424
+ },
425
+ "action": {
426
+ "delta_indices": [
427
+ 0,
428
+ 1,
429
+ 2,
430
+ 3,
431
+ 4,
432
+ 5,
433
+ 6,
434
+ 7,
435
+ 8,
436
+ 9,
437
+ 10,
438
+ 11,
439
+ 12,
440
+ 13,
441
+ 14,
442
+ 15,
443
+ 16,
444
+ 17,
445
+ 18,
446
+ 19,
447
+ 20,
448
+ 21,
449
+ 22,
450
+ 23,
451
+ 24,
452
+ 25,
453
+ 26,
454
+ 27,
455
+ 28,
456
+ 29,
457
+ 30,
458
+ 31,
459
+ 32,
460
+ 33,
461
+ 34,
462
+ 35,
463
+ 36,
464
+ 37,
465
+ 38,
466
+ 39
467
+ ],
468
+ "modality_keys": [
469
+ "left_wrist_eef",
470
+ "right_wrist_eef",
471
+ "left_hand_joints",
472
+ "right_hand_joints"
473
+ ],
474
+ "sin_cos_embedding_keys": null,
475
+ "mean_std_embedding_keys": null,
476
+ "action_configs": [
477
+ {
478
+ "rep": "RELATIVE",
479
+ "type": "EEF",
480
+ "format": "XYZ_ROT6D",
481
+ "state_key": "left_wrist_eef"
482
+ },
483
+ {
484
+ "rep": "RELATIVE",
485
+ "type": "EEF",
486
+ "format": "XYZ_ROT6D",
487
+ "state_key": "right_wrist_eef"
488
+ },
489
+ {
490
+ "rep": "ABSOLUTE",
491
+ "type": "NON_EEF",
492
+ "format": "DEFAULT",
493
+ "state_key": "left_hand_joints"
494
+ },
495
+ {
496
+ "rep": "ABSOLUTE",
497
+ "type": "NON_EEF",
498
+ "format": "DEFAULT",
499
+ "state_key": "right_hand_joints"
500
+ }
501
+ ]
502
+ },
503
+ "language": {
504
+ "delta_indices": [
505
+ 0
506
+ ],
507
+ "modality_keys": [
508
+ "annotation.human.coarse_action"
509
+ ],
510
+ "sin_cos_embedding_keys": null,
511
+ "mean_std_embedding_keys": null,
512
+ "action_configs": null
513
+ }
514
+ },
515
+ "xdof_relative_eef_relative_joint": {
516
+ "video": {
517
+ "delta_indices": [
518
+ -30,
519
+ 0
520
+ ],
521
+ "modality_keys": [
522
+ "top_camera-images-rgb_320_240",
523
+ "left_camera-images-rgb_320_240",
524
+ "right_camera-images-rgb_320_240"
525
+ ],
526
+ "sin_cos_embedding_keys": null,
527
+ "mean_std_embedding_keys": null,
528
+ "action_configs": null
529
+ },
530
+ "state": {
531
+ "delta_indices": [
532
+ 0
533
+ ],
534
+ "modality_keys": [
535
+ "left_wrist_eef",
536
+ "right_wrist_eef",
537
+ "left_gripper_pos",
538
+ "right_gripper_pos",
539
+ "left_joint_pos",
540
+ "right_joint_pos"
541
+ ],
542
+ "sin_cos_embedding_keys": null,
543
+ "mean_std_embedding_keys": null,
544
+ "action_configs": null
545
+ },
546
+ "action": {
547
+ "delta_indices": [
548
+ 0,
549
+ 1,
550
+ 2,
551
+ 3,
552
+ 4,
553
+ 5,
554
+ 6,
555
+ 7,
556
+ 8,
557
+ 9,
558
+ 10,
559
+ 11,
560
+ 12,
561
+ 13,
562
+ 14,
563
+ 15,
564
+ 16,
565
+ 17,
566
+ 18,
567
+ 19,
568
+ 20,
569
+ 21,
570
+ 22,
571
+ 23,
572
+ 24,
573
+ 25,
574
+ 26,
575
+ 27,
576
+ 28,
577
+ 29,
578
+ 30,
579
+ 31,
580
+ 32,
581
+ 33,
582
+ 34,
583
+ 35,
584
+ 36,
585
+ 37,
586
+ 38,
587
+ 39
588
+ ],
589
+ "modality_keys": [
590
+ "left_wrist_eef",
591
+ "right_wrist_eef",
592
+ "left_gripper_pos",
593
+ "right_gripper_pos",
594
+ "left_joint_pos",
595
+ "right_joint_pos"
596
+ ],
597
+ "sin_cos_embedding_keys": null,
598
+ "mean_std_embedding_keys": null,
599
+ "action_configs": [
600
+ {
601
+ "rep": "RELATIVE",
602
+ "type": "EEF",
603
+ "format": "XYZ_ROT6D",
604
+ "state_key": "left_wrist_eef"
605
+ },
606
+ {
607
+ "rep": "RELATIVE",
608
+ "type": "EEF",
609
+ "format": "XYZ_ROT6D",
610
+ "state_key": "right_wrist_eef"
611
+ },
612
+ {
613
+ "rep": "ABSOLUTE",
614
+ "type": "NON_EEF",
615
+ "format": "DEFAULT",
616
+ "state_key": "left_gripper_pos"
617
+ },
618
+ {
619
+ "rep": "ABSOLUTE",
620
+ "type": "NON_EEF",
621
+ "format": "DEFAULT",
622
+ "state_key": "right_gripper_pos"
623
+ },
624
+ {
625
+ "rep": "RELATIVE",
626
+ "type": "NON_EEF",
627
+ "format": "DEFAULT",
628
+ "state_key": "left_joint_pos"
629
+ },
630
+ {
631
+ "rep": "RELATIVE",
632
+ "type": "NON_EEF",
633
+ "format": "DEFAULT",
634
+ "state_key": "right_joint_pos"
635
+ }
636
+ ]
637
+ },
638
+ "language": {
639
+ "delta_indices": [
640
+ 0
641
+ ],
642
+ "modality_keys": [
643
+ "annotation.task"
644
+ ],
645
+ "sin_cos_embedding_keys": null,
646
+ "mean_std_embedding_keys": null,
647
+ "action_configs": null
648
+ }
649
+ },
650
+ "real_r1_pro_sharpa_relative_eef_maxinsights": {
651
+ "video": {
652
+ "delta_indices": [
653
+ -30,
654
+ 0
655
+ ],
656
+ "modality_keys": [
657
+ "ego_view_cropratio_res320x240_freq30"
658
+ ],
659
+ "sin_cos_embedding_keys": null,
660
+ "mean_std_embedding_keys": null,
661
+ "action_configs": null
662
+ },
663
+ "state": {
664
+ "delta_indices": [
665
+ 0
666
+ ],
667
+ "modality_keys": [
668
+ "left_wrist_eef",
669
+ "right_wrist_eef",
670
+ "left_hand_joints",
671
+ "right_hand_joints"
672
+ ],
673
+ "sin_cos_embedding_keys": null,
674
+ "mean_std_embedding_keys": null,
675
+ "action_configs": null
676
+ },
677
+ "action": {
678
+ "delta_indices": [
679
+ 0,
680
+ 1,
681
+ 2,
682
+ 3,
683
+ 4,
684
+ 5,
685
+ 6,
686
+ 7,
687
+ 8,
688
+ 9,
689
+ 10,
690
+ 11,
691
+ 12,
692
+ 13,
693
+ 14,
694
+ 15,
695
+ 16,
696
+ 17,
697
+ 18,
698
+ 19,
699
+ 20,
700
+ 21,
701
+ 22,
702
+ 23,
703
+ 24,
704
+ 25,
705
+ 26,
706
+ 27,
707
+ 28,
708
+ 29,
709
+ 30,
710
+ 31,
711
+ 32,
712
+ 33,
713
+ 34,
714
+ 35,
715
+ 36,
716
+ 37,
717
+ 38,
718
+ 39
719
+ ],
720
+ "modality_keys": [
721
+ "left_wrist_eef",
722
+ "right_wrist_eef",
723
+ "left_hand_joints",
724
+ "right_hand_joints"
725
+ ],
726
+ "sin_cos_embedding_keys": null,
727
+ "mean_std_embedding_keys": null,
728
+ "action_configs": [
729
+ {
730
+ "rep": "RELATIVE",
731
+ "type": "EEF",
732
+ "format": "XYZ_ROT6D",
733
+ "state_key": "left_wrist_eef"
734
+ },
735
+ {
736
+ "rep": "RELATIVE",
737
+ "type": "EEF",
738
+ "format": "XYZ_ROT6D",
739
+ "state_key": "right_wrist_eef"
740
+ },
741
+ {
742
+ "rep": "ABSOLUTE",
743
+ "type": "NON_EEF",
744
+ "format": "DEFAULT",
745
+ "state_key": "left_hand_joints"
746
+ },
747
+ {
748
+ "rep": "ABSOLUTE",
749
+ "type": "NON_EEF",
750
+ "format": "DEFAULT",
751
+ "state_key": "right_hand_joints"
752
+ }
753
+ ]
754
+ },
755
+ "language": {
756
+ "delta_indices": [
757
+ 0
758
+ ],
759
+ "modality_keys": [
760
+ "annotation.human.coarse_action"
761
+ ],
762
+ "sin_cos_embedding_keys": null,
763
+ "mean_std_embedding_keys": null,
764
+ "action_configs": null
765
+ }
766
+ },
767
+ "xdof_relative_eef_relative_joint_subtask": {
768
+ "video": {
769
+ "delta_indices": [
770
+ -30,
771
+ 0
772
+ ],
773
+ "modality_keys": [
774
+ "top_camera-images-rgb_320_240",
775
+ "left_camera-images-rgb_320_240",
776
+ "right_camera-images-rgb_320_240"
777
+ ],
778
+ "sin_cos_embedding_keys": null,
779
+ "mean_std_embedding_keys": null,
780
+ "action_configs": null
781
+ },
782
+ "state": {
783
+ "delta_indices": [
784
+ 0
785
+ ],
786
+ "modality_keys": [
787
+ "left_wrist_eef",
788
+ "right_wrist_eef",
789
+ "left_gripper_pos",
790
+ "right_gripper_pos",
791
+ "left_joint_pos",
792
+ "right_joint_pos"
793
+ ],
794
+ "sin_cos_embedding_keys": null,
795
+ "mean_std_embedding_keys": null,
796
+ "action_configs": null
797
+ },
798
+ "action": {
799
+ "delta_indices": [
800
+ 0,
801
+ 1,
802
+ 2,
803
+ 3,
804
+ 4,
805
+ 5,
806
+ 6,
807
+ 7,
808
+ 8,
809
+ 9,
810
+ 10,
811
+ 11,
812
+ 12,
813
+ 13,
814
+ 14,
815
+ 15,
816
+ 16,
817
+ 17,
818
+ 18,
819
+ 19,
820
+ 20,
821
+ 21,
822
+ 22,
823
+ 23,
824
+ 24,
825
+ 25,
826
+ 26,
827
+ 27,
828
+ 28,
829
+ 29,
830
+ 30,
831
+ 31,
832
+ 32,
833
+ 33,
834
+ 34,
835
+ 35,
836
+ 36,
837
+ 37,
838
+ 38,
839
+ 39
840
+ ],
841
+ "modality_keys": [
842
+ "left_wrist_eef",
843
+ "right_wrist_eef",
844
+ "left_gripper_pos",
845
+ "right_gripper_pos",
846
+ "left_joint_pos",
847
+ "right_joint_pos"
848
+ ],
849
+ "sin_cos_embedding_keys": null,
850
+ "mean_std_embedding_keys": null,
851
+ "action_configs": [
852
+ {
853
+ "rep": "RELATIVE",
854
+ "type": "EEF",
855
+ "format": "XYZ_ROT6D",
856
+ "state_key": "left_wrist_eef"
857
+ },
858
+ {
859
+ "rep": "RELATIVE",
860
+ "type": "EEF",
861
+ "format": "XYZ_ROT6D",
862
+ "state_key": "right_wrist_eef"
863
+ },
864
+ {
865
+ "rep": "ABSOLUTE",
866
+ "type": "NON_EEF",
867
+ "format": "DEFAULT",
868
+ "state_key": "left_gripper_pos"
869
+ },
870
+ {
871
+ "rep": "ABSOLUTE",
872
+ "type": "NON_EEF",
873
+ "format": "DEFAULT",
874
+ "state_key": "right_gripper_pos"
875
+ },
876
+ {
877
+ "rep": "RELATIVE",
878
+ "type": "NON_EEF",
879
+ "format": "DEFAULT",
880
+ "state_key": "left_joint_pos"
881
+ },
882
+ {
883
+ "rep": "RELATIVE",
884
+ "type": "NON_EEF",
885
+ "format": "DEFAULT",
886
+ "state_key": "right_joint_pos"
887
+ }
888
+ ]
889
+ },
890
+ "language": {
891
+ "delta_indices": [
892
+ 0
893
+ ],
894
+ "modality_keys": [
895
+ "annotation.sub_task"
896
+ ],
897
+ "sin_cos_embedding_keys": null,
898
+ "mean_std_embedding_keys": null,
899
+ "action_configs": null
900
+ }
901
+ },
902
+ "oxe_droid_relative_eef_relative_joint": {
903
+ "video": {
904
+ "delta_indices": [
905
+ -15,
906
+ 0
907
+ ],
908
+ "modality_keys": [
909
+ "exterior_image_1_left",
910
+ "wrist_image_left"
911
+ ],
912
+ "sin_cos_embedding_keys": null,
913
+ "mean_std_embedding_keys": null,
914
+ "action_configs": null
915
+ },
916
+ "state": {
917
+ "delta_indices": [
918
+ 0
919
+ ],
920
+ "modality_keys": [
921
+ "eef_9d",
922
+ "gripper_position",
923
+ "joint_position"
924
+ ],
925
+ "sin_cos_embedding_keys": null,
926
+ "mean_std_embedding_keys": null,
927
+ "action_configs": null
928
+ },
929
+ "action": {
930
+ "delta_indices": [
931
+ 0,
932
+ 1,
933
+ 2,
934
+ 3,
935
+ 4,
936
+ 5,
937
+ 6,
938
+ 7,
939
+ 8,
940
+ 9,
941
+ 10,
942
+ 11,
943
+ 12,
944
+ 13,
945
+ 14,
946
+ 15,
947
+ 16,
948
+ 17,
949
+ 18,
950
+ 19,
951
+ 20,
952
+ 21,
953
+ 22,
954
+ 23,
955
+ 24,
956
+ 25,
957
+ 26,
958
+ 27,
959
+ 28,
960
+ 29,
961
+ 30,
962
+ 31,
963
+ 32,
964
+ 33,
965
+ 34,
966
+ 35,
967
+ 36,
968
+ 37,
969
+ 38,
970
+ 39
971
+ ],
972
+ "modality_keys": [
973
+ "eef_9d",
974
+ "gripper_position",
975
+ "joint_position"
976
+ ],
977
+ "sin_cos_embedding_keys": null,
978
+ "mean_std_embedding_keys": null,
979
+ "action_configs": [
980
+ {
981
+ "rep": "RELATIVE",
982
+ "type": "EEF",
983
+ "format": "XYZ_ROT6D",
984
+ "state_key": "eef_9d"
985
+ },
986
+ {
987
+ "rep": "ABSOLUTE",
988
+ "type": "NON_EEF",
989
+ "format": "DEFAULT",
990
+ "state_key": "gripper_position"
991
+ },
992
+ {
993
+ "rep": "RELATIVE",
994
+ "type": "NON_EEF",
995
+ "format": "DEFAULT",
996
+ "state_key": "joint_position"
997
+ }
998
+ ]
999
+ },
1000
+ "language": {
1001
+ "delta_indices": [
1002
+ 0
1003
+ ],
1004
+ "modality_keys": [
1005
+ "annotation.language.language_instruction"
1006
+ ],
1007
+ "sin_cos_embedding_keys": null,
1008
+ "mean_std_embedding_keys": null,
1009
+ "action_configs": null
1010
+ }
1011
+ },
1012
+ "new_embodiment": {
1013
+ "video": {
1014
+ "delta_indices": [
1015
+ 0
1016
+ ],
1017
+ "modality_keys": [
1018
+ "head",
1019
+ "chest"
1020
+ ],
1021
+ "sin_cos_embedding_keys": null,
1022
+ "mean_std_embedding_keys": null,
1023
+ "action_configs": null
1024
+ },
1025
+ "state": {
1026
+ "delta_indices": [
1027
+ 0
1028
+ ],
1029
+ "modality_keys": [
1030
+ "left_leg",
1031
+ "right_leg",
1032
+ "waist",
1033
+ "left_arm",
1034
+ "right_arm",
1035
+ "head",
1036
+ "left_hand",
1037
+ "right_hand",
1038
+ "projected_gravity"
1039
+ ],
1040
+ "sin_cos_embedding_keys": null,
1041
+ "mean_std_embedding_keys": null,
1042
+ "action_configs": null
1043
+ },
1044
+ "action": {
1045
+ "delta_indices": [
1046
+ 0,
1047
+ 1,
1048
+ 2,
1049
+ 3,
1050
+ 4,
1051
+ 5,
1052
+ 6,
1053
+ 7,
1054
+ 8,
1055
+ 9,
1056
+ 10,
1057
+ 11,
1058
+ 12,
1059
+ 13,
1060
+ 14,
1061
+ 15
1062
+ ],
1063
+ "modality_keys": [
1064
+ "body_target",
1065
+ "left_hand_target",
1066
+ "right_hand_target",
1067
+ "planner"
1068
+ ],
1069
+ "sin_cos_embedding_keys": null,
1070
+ "mean_std_embedding_keys": null,
1071
+ "action_configs": [
1072
+ {
1073
+ "rep": "ABSOLUTE",
1074
+ "type": "NON_EEF",
1075
+ "format": "DEFAULT",
1076
+ "state_key": null
1077
+ },
1078
+ {
1079
+ "rep": "ABSOLUTE",
1080
+ "type": "NON_EEF",
1081
+ "format": "DEFAULT",
1082
+ "state_key": null
1083
+ },
1084
+ {
1085
+ "rep": "ABSOLUTE",
1086
+ "type": "NON_EEF",
1087
+ "format": "DEFAULT",
1088
+ "state_key": null
1089
+ },
1090
+ {
1091
+ "rep": "ABSOLUTE",
1092
+ "type": "NON_EEF",
1093
+ "format": "DEFAULT",
1094
+ "state_key": null
1095
+ }
1096
+ ]
1097
+ },
1098
+ "language": {
1099
+ "delta_indices": [
1100
+ 0
1101
+ ],
1102
+ "modality_keys": [
1103
+ "annotation.human.task_description"
1104
+ ],
1105
+ "sin_cos_embedding_keys": null,
1106
+ "mean_std_embedding_keys": null,
1107
+ "action_configs": null
1108
+ }
1109
+ }
1110
+ },
1111
+ "image_crop_size": [
1112
+ 230,
1113
+ 230
1114
+ ],
1115
+ "image_target_size": [
1116
+ 256,
1117
+ 256
1118
+ ],
1119
+ "use_albumentations": true,
1120
+ "random_rotation_angle": 0,
1121
+ "color_jitter_params": {
1122
+ "brightness": 0.3,
1123
+ "contrast": 0.4,
1124
+ "saturation": 0.5,
1125
+ "hue": 0.08
1126
+ },
1127
+ "shortest_image_edge": 256,
1128
+ "crop_fraction": 0.95,
1129
+ "letter_box_transform": false,
1130
+ "model_name": "nvidia/Cosmos-Reason2-2B",
1131
+ "model_type": "qwen",
1132
+ "formalize_language": true,
1133
+ "max_state_dim": 132,
1134
+ "max_action_dim": 132,
1135
+ "max_action_horizon": 40,
1136
+ "strict_action_padding_mask": false,
1137
+ "use_percentiles": true,
1138
+ "use_mean_std": false,
1139
+ "clip_outliers": true,
1140
+ "apply_sincos_state_encoding": false,
1141
+ "use_relative_action": true,
1142
+ "exclude_state": false,
1143
+ "state_dropout_prob": 0.2
1144
+ }
1145
+ }
a14-unscrew-bulb/v2-b512/step-002000/protocol.json ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "GR00T N1.7 (Gr00tN1d7), full fine-tune: all modules (tune_llm, tune_visual, projector, diffusion action head, VLLN)",
3
+ "task": "A14 screw in a bulb v2 (H2, cameras head + chest, action82 with the planner command)",
4
+ "task_dir": "a14-unscrew-bulb/v2-b512",
5
+ "robot": "H2 (Sharpa hands)",
6
+ "training_data": "humanoid-benchmark/a14-unscrew-bulb-expert-v2 @ 8411a6dac45d0097d6068a60307de0c2e3b621a7",
7
+ "episodes": 100,
8
+ "epoch_completed": null,
9
+ "global_step": 2000,
10
+ "augmentation": "Humanoid Benchmark online augmentation spec v2 (2026-11-09): 15 visual effects, one effect per training sample at 1.5x; proprioception aligned with the GR00T A20 recipe (2026-11-12): per-sample joint bias +-2 deg + Gaussian noise 0.25 deg (hb_aug_v2.perturb_joints), no proprio delay, no history drop; model-native image augmentation off",
11
+ "epochs": null,
12
+ "total_steps": 4000,
13
+ "save_steps": [
14
+ 2000,
15
+ 2500,
16
+ 3000,
17
+ 3500,
18
+ 4000
19
+ ],
20
+ "global_batch": 512,
21
+ "per_gpu_batch": 32,
22
+ "num_gpus": 16,
23
+ "lr": 0.0001,
24
+ "lr_schedule": "cosine, warmup_ratio 0.05",
25
+ "weight_decay": 1e-05,
26
+ "proprio_augmentation": "per-sample joint bias +-2 deg + noise 0.25 deg (hb_aug_v2.perturb_joints); no proprio delay / history drop (the GR00T A20 recipe)",
27
+ "config": "experiment_cfg/conf.yaml",
28
+ "normalization": "statistics.json",
29
+ "inference": "NVIDIA Isaac-GR00T N1.7 policy loader with embodiment_tag new_embodiment and experiment_cfg/conf.yaml; VLM nvidia/Cosmos-Reason2-2B from its own repository.",
30
+ "epochs_equivalent": 11.03,
31
+ "recipe": "full fine-tune (tune_llm, tune_visual, projector, diffusion, vlln), fixed steps, 2 nodes x 8 x 32, no accumulation",
32
+ "gradient_accumulation_steps": 1,
33
+ "num_nodes": 2,
34
+ "frames": 185638,
35
+ "trainable": "all modules (tune_llm, tune_visual, projector, diffusion action head, VLLN)",
36
+ "cameras": [
37
+ "observation.images.head",
38
+ "observation.images.chest"
39
+ ],
40
+ "split": "all 100 episodes for training",
41
+ "state_layout": "state78 = observation.state (75: legs 0-11, waist 12-14, left arm 15-21, right arm 22-28, head 29-30, hands 31-74) + observation.projected_gravity (3); no joint velocity",
42
+ "action_layout": "action82 = action.body_target (31) + action.hands (44: left 22, right 22) + action.planner (7: planner_active, walk, move_cos, move_sin, face_cos, face_sin, speed_m_s); GR00T modalities body_target / left_hand_target / right_hand_target / planner, ABSOLUTE, 16 rows per query",
43
+ "initialization": "GR00T N1.7 base (groot-n17-base); VLM nvidia/Cosmos-Reason2-2B",
44
+ "recipe_reference": "groot/a14-unscrew-bulb/step-004000 (v1, 3 cameras, action75) recipe with cameras head + chest and global batch 512 (2 nodes x 8 GPUs x 32, no gradient accumulation); everything else unchanged (H2 training guide 4.1)",
45
+ "published_steps": [
46
+ 2000,
47
+ 4000
48
+ ],
49
+ "final_loss_mean_last50": 0.02291,
50
+ "loss_curve": "loss_curve.csv",
51
+ "training_records_included": true,
52
+ "resume_state_included": false,
53
+ "files": {
54
+ "code/hb_groot/cache_cosmos_reason2.py": 2542,
55
+ "code/hb_groot/check_reason2_offline.py": 1283,
56
+ "code/hb_groot/g1_feedback76_h16.py": 1474,
57
+ "code/hb_groot/groot_chain.py": 6325,
58
+ "code/hb_groot/h2_body31_hand44_h16.py": 1395,
59
+ "code/hb_groot/h2_body31_hand44_planner7_h16.py": 1531,
60
+ "code/hb_groot/h2w3_body31_hand44_h16.py": 1563,
61
+ "code/hb_groot/hb_aug_v2.py": 7475,
62
+ "code/hb_groot/hb_launch.py": 3913,
63
+ "code/hb_groot/hb_offline_fix.py": 1138,
64
+ "code/hb_groot/prepare_groot_data.py": 12446,
65
+ "code/hb_groot/verify_groot_cpu.py": 2744,
66
+ "config.json": 2219,
67
+ "dataset_meta/hb-dataset-receipt.json": 783,
68
+ "dataset_meta/modality.json": 1646,
69
+ "embodiment_id.json": 2226,
70
+ "experiment_cfg/conf.yaml": 5925,
71
+ "experiment_cfg/config.yaml": 5893,
72
+ "experiment_cfg/dataset_statistics.json": 30898,
73
+ "experiment_cfg/final_model_config.json": 1584,
74
+ "experiment_cfg/final_processor_config.json": 2548675,
75
+ "loss_curve.csv": 21019,
76
+ "model-00001-of-00002.safetensors": 4990519232,
77
+ "model-00002-of-00002.safetensors": 1919980184,
78
+ "model.safetensors.index.json": 104985,
79
+ "processor_config.json": 27145,
80
+ "statistics.json": 4323991,
81
+ "train.log": 501726,
82
+ "trainer_state.json": 28269,
83
+ "training_args.bin": 8017,
84
+ "training_plan.json": 948,
85
+ "wandb_config.json": 58
86
+ }
87
+ }
a14-unscrew-bulb/v2-b512/step-002000/statistics.json ADDED
The diff for this file is too large to render. See raw diff
 
a14-unscrew-bulb/v2-b512/step-002000/train.log ADDED
The diff for this file is too large to render. See raw diff
 
a14-unscrew-bulb/v2-b512/step-002000/trainer_state.json ADDED
@@ -0,0 +1,1234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 0.5,
6
+ "eval_steps": 500,
7
+ "global_step": 2000,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "grad_norm": 0.45548781752586365,
14
+ "learning_rate": 4.5e-06,
15
+ "loss": 1.4311,
16
+ "step": 10
17
+ },
18
+ {
19
+ "grad_norm": 0.6781157851219177,
20
+ "learning_rate": 9.5e-06,
21
+ "loss": 1.3387,
22
+ "step": 20
23
+ },
24
+ {
25
+ "grad_norm": 0.603681743144989,
26
+ "learning_rate": 1.45e-05,
27
+ "loss": 1.2232,
28
+ "step": 30
29
+ },
30
+ {
31
+ "grad_norm": 0.5604022741317749,
32
+ "learning_rate": 1.9500000000000003e-05,
33
+ "loss": 1.1367,
34
+ "step": 40
35
+ },
36
+ {
37
+ "grad_norm": 1.2952625751495361,
38
+ "learning_rate": 2.45e-05,
39
+ "loss": 1.0905,
40
+ "step": 50
41
+ },
42
+ {
43
+ "grad_norm": 0.6026202440261841,
44
+ "learning_rate": 2.95e-05,
45
+ "loss": 1.0696,
46
+ "step": 60
47
+ },
48
+ {
49
+ "grad_norm": 0.3096122145652771,
50
+ "learning_rate": 3.45e-05,
51
+ "loss": 1.0596,
52
+ "step": 70
53
+ },
54
+ {
55
+ "grad_norm": 0.2874182462692261,
56
+ "learning_rate": 3.9500000000000005e-05,
57
+ "loss": 1.0537,
58
+ "step": 80
59
+ },
60
+ {
61
+ "grad_norm": 0.2818102240562439,
62
+ "learning_rate": 4.4500000000000004e-05,
63
+ "loss": 1.0422,
64
+ "step": 90
65
+ },
66
+ {
67
+ "grad_norm": 0.4246894121170044,
68
+ "learning_rate": 4.9500000000000004e-05,
69
+ "loss": 1.0423,
70
+ "step": 100
71
+ },
72
+ {
73
+ "grad_norm": 0.5688769817352295,
74
+ "learning_rate": 5.45e-05,
75
+ "loss": 1.0333,
76
+ "step": 110
77
+ },
78
+ {
79
+ "grad_norm": 0.5930460691452026,
80
+ "learning_rate": 5.95e-05,
81
+ "loss": 0.9992,
82
+ "step": 120
83
+ },
84
+ {
85
+ "grad_norm": 0.5382891893386841,
86
+ "learning_rate": 6.450000000000001e-05,
87
+ "loss": 0.9417,
88
+ "step": 130
89
+ },
90
+ {
91
+ "grad_norm": 0.7342086434364319,
92
+ "learning_rate": 6.95e-05,
93
+ "loss": 0.8895,
94
+ "step": 140
95
+ },
96
+ {
97
+ "grad_norm": 0.8100906014442444,
98
+ "learning_rate": 7.450000000000001e-05,
99
+ "loss": 0.8431,
100
+ "step": 150
101
+ },
102
+ {
103
+ "grad_norm": 0.6786221861839294,
104
+ "learning_rate": 7.950000000000001e-05,
105
+ "loss": 0.8041,
106
+ "step": 160
107
+ },
108
+ {
109
+ "grad_norm": 0.8967251777648926,
110
+ "learning_rate": 8.450000000000001e-05,
111
+ "loss": 0.7719,
112
+ "step": 170
113
+ },
114
+ {
115
+ "grad_norm": 0.7777813673019409,
116
+ "learning_rate": 8.950000000000001e-05,
117
+ "loss": 0.7539,
118
+ "step": 180
119
+ },
120
+ {
121
+ "grad_norm": 0.8444279432296753,
122
+ "learning_rate": 9.449999999999999e-05,
123
+ "loss": 0.7309,
124
+ "step": 190
125
+ },
126
+ {
127
+ "grad_norm": 0.7900912165641785,
128
+ "learning_rate": 9.95e-05,
129
+ "loss": 0.718,
130
+ "step": 200
131
+ },
132
+ {
133
+ "grad_norm": 0.6578666567802429,
134
+ "learning_rate": 9.999861593790126e-05,
135
+ "loss": 0.7104,
136
+ "step": 210
137
+ },
138
+ {
139
+ "grad_norm": 0.7282286882400513,
140
+ "learning_rate": 9.999383162408304e-05,
141
+ "loss": 0.7047,
142
+ "step": 220
143
+ },
144
+ {
145
+ "grad_norm": 0.7857356667518616,
146
+ "learning_rate": 9.998563029828259e-05,
147
+ "loss": 0.7024,
148
+ "step": 230
149
+ },
150
+ {
151
+ "grad_norm": 0.7012117505073547,
152
+ "learning_rate": 9.997401252104962e-05,
153
+ "loss": 0.6952,
154
+ "step": 240
155
+ },
156
+ {
157
+ "grad_norm": 0.6767271161079407,
158
+ "learning_rate": 9.995897908644378e-05,
159
+ "loss": 0.6898,
160
+ "step": 250
161
+ },
162
+ {
163
+ "grad_norm": 0.8207566142082214,
164
+ "learning_rate": 9.994053102198034e-05,
165
+ "loss": 0.6809,
166
+ "step": 260
167
+ },
168
+ {
169
+ "grad_norm": 0.9090561866760254,
170
+ "learning_rate": 9.991866958856003e-05,
171
+ "loss": 0.6699,
172
+ "step": 270
173
+ },
174
+ {
175
+ "grad_norm": 0.833543598651886,
176
+ "learning_rate": 9.989339628038276e-05,
177
+ "loss": 0.6438,
178
+ "step": 280
179
+ },
180
+ {
181
+ "grad_norm": 0.7834540605545044,
182
+ "learning_rate": 9.98647128248456e-05,
183
+ "loss": 0.6349,
184
+ "step": 290
185
+ },
186
+ {
187
+ "grad_norm": 0.9627916216850281,
188
+ "learning_rate": 9.98326211824246e-05,
189
+ "loss": 0.6051,
190
+ "step": 300
191
+ },
192
+ {
193
+ "grad_norm": 0.7982354760169983,
194
+ "learning_rate": 9.979712354654091e-05,
195
+ "loss": 0.5885,
196
+ "step": 310
197
+ },
198
+ {
199
+ "grad_norm": 0.7507064342498779,
200
+ "learning_rate": 9.975822234341079e-05,
201
+ "loss": 0.5621,
202
+ "step": 320
203
+ },
204
+ {
205
+ "grad_norm": 1.0718227624893188,
206
+ "learning_rate": 9.97159202318798e-05,
207
+ "loss": 0.5466,
208
+ "step": 330
209
+ },
210
+ {
211
+ "grad_norm": 1.0223731994628906,
212
+ "learning_rate": 9.967022010324105e-05,
213
+ "loss": 0.5236,
214
+ "step": 340
215
+ },
216
+ {
217
+ "grad_norm": 0.8068194389343262,
218
+ "learning_rate": 9.962112508103765e-05,
219
+ "loss": 0.5016,
220
+ "step": 350
221
+ },
222
+ {
223
+ "grad_norm": 1.0485377311706543,
224
+ "learning_rate": 9.956863852084914e-05,
225
+ "loss": 0.484,
226
+ "step": 360
227
+ },
228
+ {
229
+ "grad_norm": 0.8841873407363892,
230
+ "learning_rate": 9.951276401006221e-05,
231
+ "loss": 0.4736,
232
+ "step": 370
233
+ },
234
+ {
235
+ "grad_norm": 1.0380600690841675,
236
+ "learning_rate": 9.945350536762543e-05,
237
+ "loss": 0.4506,
238
+ "step": 380
239
+ },
240
+ {
241
+ "grad_norm": 1.1374574899673462,
242
+ "learning_rate": 9.939086664378829e-05,
243
+ "loss": 0.4317,
244
+ "step": 390
245
+ },
246
+ {
247
+ "grad_norm": 1.1841020584106445,
248
+ "learning_rate": 9.932485211982437e-05,
249
+ "loss": 0.4166,
250
+ "step": 400
251
+ },
252
+ {
253
+ "grad_norm": 1.1365927457809448,
254
+ "learning_rate": 9.92554663077387e-05,
255
+ "loss": 0.4057,
256
+ "step": 410
257
+ },
258
+ {
259
+ "grad_norm": 1.2993481159210205,
260
+ "learning_rate": 9.918271394995935e-05,
261
+ "loss": 0.3918,
262
+ "step": 420
263
+ },
264
+ {
265
+ "grad_norm": 1.1606972217559814,
266
+ "learning_rate": 9.910660001901335e-05,
267
+ "loss": 0.3842,
268
+ "step": 430
269
+ },
270
+ {
271
+ "grad_norm": 1.272127628326416,
272
+ "learning_rate": 9.902712971718675e-05,
273
+ "loss": 0.3623,
274
+ "step": 440
275
+ },
276
+ {
277
+ "grad_norm": 1.0958130359649658,
278
+ "learning_rate": 9.894430847616915e-05,
279
+ "loss": 0.3492,
280
+ "step": 450
281
+ },
282
+ {
283
+ "grad_norm": 1.047905683517456,
284
+ "learning_rate": 9.885814195668232e-05,
285
+ "loss": 0.3361,
286
+ "step": 460
287
+ },
288
+ {
289
+ "grad_norm": 1.2776614427566528,
290
+ "learning_rate": 9.876863604809344e-05,
291
+ "loss": 0.3306,
292
+ "step": 470
293
+ },
294
+ {
295
+ "grad_norm": 1.2377774715423584,
296
+ "learning_rate": 9.867579686801245e-05,
297
+ "loss": 0.3091,
298
+ "step": 480
299
+ },
300
+ {
301
+ "grad_norm": 1.2725355625152588,
302
+ "learning_rate": 9.8579630761874e-05,
303
+ "loss": 0.2942,
304
+ "step": 490
305
+ },
306
+ {
307
+ "grad_norm": 1.1680372953414917,
308
+ "learning_rate": 9.848014430250367e-05,
309
+ "loss": 0.2801,
310
+ "step": 500
311
+ },
312
+ {
313
+ "grad_norm": 1.1904488801956177,
314
+ "learning_rate": 9.837734428966885e-05,
315
+ "loss": 0.2755,
316
+ "step": 510
317
+ },
318
+ {
319
+ "grad_norm": 1.2455122470855713,
320
+ "learning_rate": 9.827123774961383e-05,
321
+ "loss": 0.2507,
322
+ "step": 520
323
+ },
324
+ {
325
+ "grad_norm": 1.3304517269134521,
326
+ "learning_rate": 9.816183193457968e-05,
327
+ "loss": 0.249,
328
+ "step": 530
329
+ },
330
+ {
331
+ "grad_norm": 1.105523943901062,
332
+ "learning_rate": 9.804913432230856e-05,
333
+ "loss": 0.2308,
334
+ "step": 540
335
+ },
336
+ {
337
+ "grad_norm": 1.0879801511764526,
338
+ "learning_rate": 9.793315261553252e-05,
339
+ "loss": 0.2225,
340
+ "step": 550
341
+ },
342
+ {
343
+ "grad_norm": 1.3669681549072266,
344
+ "learning_rate": 9.781389474144717e-05,
345
+ "loss": 0.2076,
346
+ "step": 560
347
+ },
348
+ {
349
+ "grad_norm": 1.2130101919174194,
350
+ "learning_rate": 9.76913688511698e-05,
351
+ "loss": 0.193,
352
+ "step": 570
353
+ },
354
+ {
355
+ "grad_norm": 1.3893616199493408,
356
+ "learning_rate": 9.756558331918227e-05,
357
+ "loss": 0.1868,
358
+ "step": 580
359
+ },
360
+ {
361
+ "grad_norm": 1.3393754959106445,
362
+ "learning_rate": 9.743654674275855e-05,
363
+ "loss": 0.1777,
364
+ "step": 590
365
+ },
366
+ {
367
+ "grad_norm": 1.2126128673553467,
368
+ "learning_rate": 9.730426794137727e-05,
369
+ "loss": 0.1676,
370
+ "step": 600
371
+ },
372
+ {
373
+ "grad_norm": 1.4670941829681396,
374
+ "learning_rate": 9.716875595611879e-05,
375
+ "loss": 0.1637,
376
+ "step": 610
377
+ },
378
+ {
379
+ "grad_norm": 1.126127004623413,
380
+ "learning_rate": 9.703002004904729e-05,
381
+ "loss": 0.1378,
382
+ "step": 620
383
+ },
384
+ {
385
+ "grad_norm": 1.18602454662323,
386
+ "learning_rate": 9.688806970257773e-05,
387
+ "loss": 0.1278,
388
+ "step": 630
389
+ },
390
+ {
391
+ "grad_norm": 1.3860094547271729,
392
+ "learning_rate": 9.674291461882774e-05,
393
+ "loss": 0.1231,
394
+ "step": 640
395
+ },
396
+ {
397
+ "grad_norm": 1.3980443477630615,
398
+ "learning_rate": 9.659456471895445e-05,
399
+ "loss": 0.1235,
400
+ "step": 650
401
+ },
402
+ {
403
+ "grad_norm": 1.209280252456665,
404
+ "learning_rate": 9.644303014247648e-05,
405
+ "loss": 0.1123,
406
+ "step": 660
407
+ },
408
+ {
409
+ "grad_norm": 1.2374107837677002,
410
+ "learning_rate": 9.628832124658085e-05,
411
+ "loss": 0.1269,
412
+ "step": 670
413
+ },
414
+ {
415
+ "grad_norm": 1.175325632095337,
416
+ "learning_rate": 9.613044860541507e-05,
417
+ "loss": 0.1101,
418
+ "step": 680
419
+ },
420
+ {
421
+ "grad_norm": 1.1790177822113037,
422
+ "learning_rate": 9.596942300936445e-05,
423
+ "loss": 0.1031,
424
+ "step": 690
425
+ },
426
+ {
427
+ "grad_norm": 1.1817500591278076,
428
+ "learning_rate": 9.580525546431459e-05,
429
+ "loss": 0.1052,
430
+ "step": 700
431
+ },
432
+ {
433
+ "grad_norm": 1.258915901184082,
434
+ "learning_rate": 9.563795719089911e-05,
435
+ "loss": 0.1116,
436
+ "step": 710
437
+ },
438
+ {
439
+ "grad_norm": 1.2639306783676147,
440
+ "learning_rate": 9.546753962373281e-05,
441
+ "loss": 0.0937,
442
+ "step": 720
443
+ },
444
+ {
445
+ "grad_norm": 1.1046220064163208,
446
+ "learning_rate": 9.529401441062997e-05,
447
+ "loss": 0.0928,
448
+ "step": 730
449
+ },
450
+ {
451
+ "grad_norm": 1.2190810441970825,
452
+ "learning_rate": 9.511739341180842e-05,
453
+ "loss": 0.0788,
454
+ "step": 740
455
+ },
456
+ {
457
+ "grad_norm": 1.2571676969528198,
458
+ "learning_rate": 9.493768869907886e-05,
459
+ "loss": 0.0894,
460
+ "step": 750
461
+ },
462
+ {
463
+ "grad_norm": 1.0401405096054077,
464
+ "learning_rate": 9.475491255501968e-05,
465
+ "loss": 0.0807,
466
+ "step": 760
467
+ },
468
+ {
469
+ "grad_norm": 1.1861028671264648,
470
+ "learning_rate": 9.456907747213748e-05,
471
+ "loss": 0.0829,
472
+ "step": 770
473
+ },
474
+ {
475
+ "grad_norm": 1.1651418209075928,
476
+ "learning_rate": 9.438019615201336e-05,
477
+ "loss": 0.0843,
478
+ "step": 780
479
+ },
480
+ {
481
+ "grad_norm": 0.9502962827682495,
482
+ "learning_rate": 9.418828150443469e-05,
483
+ "loss": 0.0788,
484
+ "step": 790
485
+ },
486
+ {
487
+ "grad_norm": 1.1027251482009888,
488
+ "learning_rate": 9.399334664651262e-05,
489
+ "loss": 0.0842,
490
+ "step": 800
491
+ },
492
+ {
493
+ "grad_norm": 1.1836992502212524,
494
+ "learning_rate": 9.379540490178581e-05,
495
+ "loss": 0.088,
496
+ "step": 810
497
+ },
498
+ {
499
+ "grad_norm": 1.1526118516921997,
500
+ "learning_rate": 9.359446979930955e-05,
501
+ "loss": 0.0818,
502
+ "step": 820
503
+ },
504
+ {
505
+ "grad_norm": 1.2146070003509521,
506
+ "learning_rate": 9.33905550727312e-05,
507
+ "loss": 0.0822,
508
+ "step": 830
509
+ },
510
+ {
511
+ "grad_norm": 1.1088842153549194,
512
+ "learning_rate": 9.318367465935142e-05,
513
+ "loss": 0.0762,
514
+ "step": 840
515
+ },
516
+ {
517
+ "grad_norm": 1.0168155431747437,
518
+ "learning_rate": 9.29738426991717e-05,
519
+ "loss": 0.0873,
520
+ "step": 850
521
+ },
522
+ {
523
+ "grad_norm": 1.0673962831497192,
524
+ "learning_rate": 9.276107353392774e-05,
525
+ "loss": 0.0771,
526
+ "step": 860
527
+ },
528
+ {
529
+ "grad_norm": 1.0738420486450195,
530
+ "learning_rate": 9.254538170610938e-05,
531
+ "loss": 0.0773,
532
+ "step": 870
533
+ },
534
+ {
535
+ "grad_norm": 1.0947306156158447,
536
+ "learning_rate": 9.232678195796654e-05,
537
+ "loss": 0.0721,
538
+ "step": 880
539
+ },
540
+ {
541
+ "grad_norm": 1.12360417842865,
542
+ "learning_rate": 9.210528923050164e-05,
543
+ "loss": 0.0691,
544
+ "step": 890
545
+ },
546
+ {
547
+ "grad_norm": 1.0733106136322021,
548
+ "learning_rate": 9.188091866244834e-05,
549
+ "loss": 0.0812,
550
+ "step": 900
551
+ },
552
+ {
553
+ "grad_norm": 1.0893988609313965,
554
+ "learning_rate": 9.165368558923695e-05,
555
+ "loss": 0.0654,
556
+ "step": 910
557
+ },
558
+ {
559
+ "grad_norm": 1.1471936702728271,
560
+ "learning_rate": 9.142360554194618e-05,
561
+ "loss": 0.0753,
562
+ "step": 920
563
+ },
564
+ {
565
+ "grad_norm": 1.1242029666900635,
566
+ "learning_rate": 9.119069424624163e-05,
567
+ "loss": 0.0682,
568
+ "step": 930
569
+ },
570
+ {
571
+ "grad_norm": 1.0201963186264038,
572
+ "learning_rate": 9.0954967621301e-05,
573
+ "loss": 0.0655,
574
+ "step": 940
575
+ },
576
+ {
577
+ "grad_norm": 1.1408215761184692,
578
+ "learning_rate": 9.071644177872594e-05,
579
+ "loss": 0.0763,
580
+ "step": 950
581
+ },
582
+ {
583
+ "grad_norm": 1.054488182067871,
584
+ "learning_rate": 9.047513302144095e-05,
585
+ "loss": 0.0622,
586
+ "step": 960
587
+ },
588
+ {
589
+ "grad_norm": 1.0666213035583496,
590
+ "learning_rate": 9.023105784257906e-05,
591
+ "loss": 0.0644,
592
+ "step": 970
593
+ },
594
+ {
595
+ "grad_norm": 1.07891047000885,
596
+ "learning_rate": 8.998423292435454e-05,
597
+ "loss": 0.0695,
598
+ "step": 980
599
+ },
600
+ {
601
+ "grad_norm": 1.0655622482299805,
602
+ "learning_rate": 8.973467513692265e-05,
603
+ "loss": 0.0704,
604
+ "step": 990
605
+ },
606
+ {
607
+ "grad_norm": 1.116550326347351,
608
+ "learning_rate": 8.94824015372267e-05,
609
+ "loss": 0.0674,
610
+ "step": 1000
611
+ },
612
+ {
613
+ "grad_norm": 0.8838066458702087,
614
+ "learning_rate": 8.922742936783207e-05,
615
+ "loss": 0.0631,
616
+ "step": 1010
617
+ },
618
+ {
619
+ "grad_norm": 1.086512565612793,
620
+ "learning_rate": 8.896977605574788e-05,
621
+ "loss": 0.0665,
622
+ "step": 1020
623
+ },
624
+ {
625
+ "grad_norm": 1.0114772319793701,
626
+ "learning_rate": 8.870945921123576e-05,
627
+ "loss": 0.0653,
628
+ "step": 1030
629
+ },
630
+ {
631
+ "grad_norm": 1.0343341827392578,
632
+ "learning_rate": 8.844649662660624e-05,
633
+ "loss": 0.067,
634
+ "step": 1040
635
+ },
636
+ {
637
+ "grad_norm": 0.9564461708068848,
638
+ "learning_rate": 8.818090627500266e-05,
639
+ "loss": 0.0594,
640
+ "step": 1050
641
+ },
642
+ {
643
+ "grad_norm": 0.8460029363632202,
644
+ "learning_rate": 8.791270630917275e-05,
645
+ "loss": 0.0734,
646
+ "step": 1060
647
+ },
648
+ {
649
+ "grad_norm": 1.006821870803833,
650
+ "learning_rate": 8.764191506022795e-05,
651
+ "loss": 0.0675,
652
+ "step": 1070
653
+ },
654
+ {
655
+ "grad_norm": 0.9695994257926941,
656
+ "learning_rate": 8.736855103639037e-05,
657
+ "loss": 0.0678,
658
+ "step": 1080
659
+ },
660
+ {
661
+ "grad_norm": 1.0134460926055908,
662
+ "learning_rate": 8.709263292172794e-05,
663
+ "loss": 0.0532,
664
+ "step": 1090
665
+ },
666
+ {
667
+ "grad_norm": 1.0870625972747803,
668
+ "learning_rate": 8.681417957487729e-05,
669
+ "loss": 0.0612,
670
+ "step": 1100
671
+ },
672
+ {
673
+ "grad_norm": 1.0378804206848145,
674
+ "learning_rate": 8.653321002775478e-05,
675
+ "loss": 0.0664,
676
+ "step": 1110
677
+ },
678
+ {
679
+ "grad_norm": 1.0451186895370483,
680
+ "learning_rate": 8.624974348425574e-05,
681
+ "loss": 0.0661,
682
+ "step": 1120
683
+ },
684
+ {
685
+ "grad_norm": 1.0146639347076416,
686
+ "learning_rate": 8.596379931894188e-05,
687
+ "loss": 0.0711,
688
+ "step": 1130
689
+ },
690
+ {
691
+ "grad_norm": 0.9299578070640564,
692
+ "learning_rate": 8.567539707571703e-05,
693
+ "loss": 0.0682,
694
+ "step": 1140
695
+ },
696
+ {
697
+ "grad_norm": 1.147046446800232,
698
+ "learning_rate": 8.538455646649146e-05,
699
+ "loss": 0.0652,
700
+ "step": 1150
701
+ },
702
+ {
703
+ "grad_norm": 1.0006110668182373,
704
+ "learning_rate": 8.509129736983446e-05,
705
+ "loss": 0.0623,
706
+ "step": 1160
707
+ },
708
+ {
709
+ "grad_norm": 0.9463521838188171,
710
+ "learning_rate": 8.479563982961571e-05,
711
+ "loss": 0.0587,
712
+ "step": 1170
713
+ },
714
+ {
715
+ "grad_norm": 1.073960781097412,
716
+ "learning_rate": 8.449760405363539e-05,
717
+ "loss": 0.0605,
718
+ "step": 1180
719
+ },
720
+ {
721
+ "grad_norm": 1.012939214706421,
722
+ "learning_rate": 8.419721041224287e-05,
723
+ "loss": 0.0629,
724
+ "step": 1190
725
+ },
726
+ {
727
+ "grad_norm": 0.8890576958656311,
728
+ "learning_rate": 8.389447943694451e-05,
729
+ "loss": 0.065,
730
+ "step": 1200
731
+ },
732
+ {
733
+ "grad_norm": 0.9021549820899963,
734
+ "learning_rate": 8.358943181900032e-05,
735
+ "loss": 0.0614,
736
+ "step": 1210
737
+ },
738
+ {
739
+ "grad_norm": 0.9169679880142212,
740
+ "learning_rate": 8.328208840800981e-05,
741
+ "loss": 0.0556,
742
+ "step": 1220
743
+ },
744
+ {
745
+ "grad_norm": 1.0166311264038086,
746
+ "learning_rate": 8.297247021048686e-05,
747
+ "loss": 0.0624,
748
+ "step": 1230
749
+ },
750
+ {
751
+ "grad_norm": 0.9880717396736145,
752
+ "learning_rate": 8.266059838842396e-05,
753
+ "loss": 0.0556,
754
+ "step": 1240
755
+ },
756
+ {
757
+ "grad_norm": 1.0422134399414062,
758
+ "learning_rate": 8.23464942578459e-05,
759
+ "loss": 0.062,
760
+ "step": 1250
761
+ },
762
+ {
763
+ "grad_norm": 0.9401381611824036,
764
+ "learning_rate": 8.203017928735277e-05,
765
+ "loss": 0.0549,
766
+ "step": 1260
767
+ },
768
+ {
769
+ "grad_norm": 1.0414115190505981,
770
+ "learning_rate": 8.17116750966526e-05,
771
+ "loss": 0.0602,
772
+ "step": 1270
773
+ },
774
+ {
775
+ "grad_norm": 0.9322880506515503,
776
+ "learning_rate": 8.139100345508377e-05,
777
+ "loss": 0.0582,
778
+ "step": 1280
779
+ },
780
+ {
781
+ "grad_norm": 1.06840980052948,
782
+ "learning_rate": 8.106818628012697e-05,
783
+ "loss": 0.0602,
784
+ "step": 1290
785
+ },
786
+ {
787
+ "grad_norm": 1.0114227533340454,
788
+ "learning_rate": 8.074324563590736e-05,
789
+ "loss": 0.0564,
790
+ "step": 1300
791
+ },
792
+ {
793
+ "grad_norm": 0.93513423204422,
794
+ "learning_rate": 8.041620373168628e-05,
795
+ "loss": 0.0534,
796
+ "step": 1310
797
+ },
798
+ {
799
+ "grad_norm": 0.9780723452568054,
800
+ "learning_rate": 8.008708292034349e-05,
801
+ "loss": 0.0585,
802
+ "step": 1320
803
+ },
804
+ {
805
+ "grad_norm": 0.8210321068763733,
806
+ "learning_rate": 7.975590569684925e-05,
807
+ "loss": 0.0513,
808
+ "step": 1330
809
+ },
810
+ {
811
+ "grad_norm": 0.9770547151565552,
812
+ "learning_rate": 7.942269469672687e-05,
813
+ "loss": 0.0541,
814
+ "step": 1340
815
+ },
816
+ {
817
+ "grad_norm": 1.064526081085205,
818
+ "learning_rate": 7.908747269450558e-05,
819
+ "loss": 0.0487,
820
+ "step": 1350
821
+ },
822
+ {
823
+ "grad_norm": 0.8699714541435242,
824
+ "learning_rate": 7.875026260216393e-05,
825
+ "loss": 0.0527,
826
+ "step": 1360
827
+ },
828
+ {
829
+ "grad_norm": 0.8412991762161255,
830
+ "learning_rate": 7.841108746756382e-05,
831
+ "loss": 0.0439,
832
+ "step": 1370
833
+ },
834
+ {
835
+ "grad_norm": 0.8587125539779663,
836
+ "learning_rate": 7.806997047287516e-05,
837
+ "loss": 0.0492,
838
+ "step": 1380
839
+ },
840
+ {
841
+ "grad_norm": 1.053528070449829,
842
+ "learning_rate": 7.772693493299138e-05,
843
+ "loss": 0.0457,
844
+ "step": 1390
845
+ },
846
+ {
847
+ "grad_norm": 0.9603511691093445,
848
+ "learning_rate": 7.7382004293936e-05,
849
+ "loss": 0.0549,
850
+ "step": 1400
851
+ },
852
+ {
853
+ "grad_norm": 0.9119125008583069,
854
+ "learning_rate": 7.703520213126e-05,
855
+ "loss": 0.0456,
856
+ "step": 1410
857
+ },
858
+ {
859
+ "grad_norm": 0.8282575607299805,
860
+ "learning_rate": 7.66865521484305e-05,
861
+ "loss": 0.0537,
862
+ "step": 1420
863
+ },
864
+ {
865
+ "grad_norm": 0.8297218084335327,
866
+ "learning_rate": 7.633607817521074e-05,
867
+ "loss": 0.0504,
868
+ "step": 1430
869
+ },
870
+ {
871
+ "grad_norm": 0.9777206778526306,
872
+ "learning_rate": 7.598380416603119e-05,
873
+ "loss": 0.051,
874
+ "step": 1440
875
+ },
876
+ {
877
+ "grad_norm": 0.8542219400405884,
878
+ "learning_rate": 7.562975419835247e-05,
879
+ "loss": 0.0536,
880
+ "step": 1450
881
+ },
882
+ {
883
+ "grad_norm": 0.8440294861793518,
884
+ "learning_rate": 7.527395247101956e-05,
885
+ "loss": 0.0504,
886
+ "step": 1460
887
+ },
888
+ {
889
+ "grad_norm": 0.7946062684059143,
890
+ "learning_rate": 7.491642330260789e-05,
891
+ "loss": 0.0502,
892
+ "step": 1470
893
+ },
894
+ {
895
+ "grad_norm": 0.9157381057739258,
896
+ "learning_rate": 7.45571911297612e-05,
897
+ "loss": 0.0563,
898
+ "step": 1480
899
+ },
900
+ {
901
+ "grad_norm": 0.8264444470405579,
902
+ "learning_rate": 7.419628050552131e-05,
903
+ "loss": 0.046,
904
+ "step": 1490
905
+ },
906
+ {
907
+ "grad_norm": 0.9470329880714417,
908
+ "learning_rate": 7.383371609764999e-05,
909
+ "loss": 0.0591,
910
+ "step": 1500
911
+ },
912
+ {
913
+ "grad_norm": 0.7743332982063293,
914
+ "learning_rate": 7.346952268694288e-05,
915
+ "loss": 0.0482,
916
+ "step": 1510
917
+ },
918
+ {
919
+ "grad_norm": 1.101076364517212,
920
+ "learning_rate": 7.310372516553585e-05,
921
+ "loss": 0.0495,
922
+ "step": 1520
923
+ },
924
+ {
925
+ "grad_norm": 0.8662927746772766,
926
+ "learning_rate": 7.273634853520356e-05,
927
+ "loss": 0.0497,
928
+ "step": 1530
929
+ },
930
+ {
931
+ "grad_norm": 0.8429809808731079,
932
+ "learning_rate": 7.236741790565072e-05,
933
+ "loss": 0.0563,
934
+ "step": 1540
935
+ },
936
+ {
937
+ "grad_norm": 0.9624887704849243,
938
+ "learning_rate": 7.199695849279576e-05,
939
+ "loss": 0.052,
940
+ "step": 1550
941
+ },
942
+ {
943
+ "grad_norm": 0.887871265411377,
944
+ "learning_rate": 7.162499561704747e-05,
945
+ "loss": 0.048,
946
+ "step": 1560
947
+ },
948
+ {
949
+ "grad_norm": 0.9352372288703918,
950
+ "learning_rate": 7.125155470157429e-05,
951
+ "loss": 0.0459,
952
+ "step": 1570
953
+ },
954
+ {
955
+ "grad_norm": 0.8238629698753357,
956
+ "learning_rate": 7.087666127056675e-05,
957
+ "loss": 0.044,
958
+ "step": 1580
959
+ },
960
+ {
961
+ "grad_norm": 0.8751165866851807,
962
+ "learning_rate": 7.050034094749286e-05,
963
+ "loss": 0.0476,
964
+ "step": 1590
965
+ },
966
+ {
967
+ "grad_norm": 0.8804195523262024,
968
+ "learning_rate": 7.012261945334683e-05,
969
+ "loss": 0.0461,
970
+ "step": 1600
971
+ },
972
+ {
973
+ "grad_norm": 0.7688097357749939,
974
+ "learning_rate": 6.974352260489103e-05,
975
+ "loss": 0.0402,
976
+ "step": 1610
977
+ },
978
+ {
979
+ "grad_norm": 0.6759632229804993,
980
+ "learning_rate": 6.936307631289148e-05,
981
+ "loss": 0.0401,
982
+ "step": 1620
983
+ },
984
+ {
985
+ "grad_norm": 0.8520809412002563,
986
+ "learning_rate": 6.898130658034685e-05,
987
+ "loss": 0.0407,
988
+ "step": 1630
989
+ },
990
+ {
991
+ "grad_norm": 1.0360755920410156,
992
+ "learning_rate": 6.859823950071127e-05,
993
+ "loss": 0.046,
994
+ "step": 1640
995
+ },
996
+ {
997
+ "grad_norm": 0.9505870938301086,
998
+ "learning_rate": 6.821390125611078e-05,
999
+ "loss": 0.0504,
1000
+ "step": 1650
1001
+ },
1002
+ {
1003
+ "grad_norm": 0.7323496341705322,
1004
+ "learning_rate": 6.782831811555385e-05,
1005
+ "loss": 0.037,
1006
+ "step": 1660
1007
+ },
1008
+ {
1009
+ "grad_norm": 0.8076654076576233,
1010
+ "learning_rate": 6.744151643313597e-05,
1011
+ "loss": 0.0441,
1012
+ "step": 1670
1013
+ },
1014
+ {
1015
+ "grad_norm": 0.8517192602157593,
1016
+ "learning_rate": 6.705352264623828e-05,
1017
+ "loss": 0.048,
1018
+ "step": 1680
1019
+ },
1020
+ {
1021
+ "grad_norm": 0.8317267894744873,
1022
+ "learning_rate": 6.666436327372078e-05,
1023
+ "loss": 0.048,
1024
+ "step": 1690
1025
+ },
1026
+ {
1027
+ "grad_norm": 0.9000750184059143,
1028
+ "learning_rate": 6.62740649141096e-05,
1029
+ "loss": 0.0439,
1030
+ "step": 1700
1031
+ },
1032
+ {
1033
+ "grad_norm": 0.8319242000579834,
1034
+ "learning_rate": 6.588265424377919e-05,
1035
+ "loss": 0.043,
1036
+ "step": 1710
1037
+ },
1038
+ {
1039
+ "grad_norm": 0.6666507720947266,
1040
+ "learning_rate": 6.549015801512895e-05,
1041
+ "loss": 0.0473,
1042
+ "step": 1720
1043
+ },
1044
+ {
1045
+ "grad_norm": 0.8333040475845337,
1046
+ "learning_rate": 6.509660305475468e-05,
1047
+ "loss": 0.045,
1048
+ "step": 1730
1049
+ },
1050
+ {
1051
+ "grad_norm": 0.8183309435844421,
1052
+ "learning_rate": 6.47020162616152e-05,
1053
+ "loss": 0.0396,
1054
+ "step": 1740
1055
+ },
1056
+ {
1057
+ "grad_norm": 0.9074743390083313,
1058
+ "learning_rate": 6.430642460519365e-05,
1059
+ "loss": 0.0428,
1060
+ "step": 1750
1061
+ },
1062
+ {
1063
+ "grad_norm": 0.8012915253639221,
1064
+ "learning_rate": 6.390985512365426e-05,
1065
+ "loss": 0.0477,
1066
+ "step": 1760
1067
+ },
1068
+ {
1069
+ "grad_norm": 0.7697868943214417,
1070
+ "learning_rate": 6.351233492199431e-05,
1071
+ "loss": 0.044,
1072
+ "step": 1770
1073
+ },
1074
+ {
1075
+ "grad_norm": 0.7668655514717102,
1076
+ "learning_rate": 6.311389117019155e-05,
1077
+ "loss": 0.0422,
1078
+ "step": 1780
1079
+ },
1080
+ {
1081
+ "grad_norm": 0.8269855380058289,
1082
+ "learning_rate": 6.271455110134713e-05,
1083
+ "loss": 0.0476,
1084
+ "step": 1790
1085
+ },
1086
+ {
1087
+ "grad_norm": 0.8340345025062561,
1088
+ "learning_rate": 6.231434200982428e-05,
1089
+ "loss": 0.0429,
1090
+ "step": 1800
1091
+ },
1092
+ {
1093
+ "grad_norm": 0.8056420683860779,
1094
+ "learning_rate": 6.191329124938285e-05,
1095
+ "loss": 0.0411,
1096
+ "step": 1810
1097
+ },
1098
+ {
1099
+ "grad_norm": 0.7833528518676758,
1100
+ "learning_rate": 6.15114262313095e-05,
1101
+ "loss": 0.0375,
1102
+ "step": 1820
1103
+ },
1104
+ {
1105
+ "grad_norm": 0.7766333818435669,
1106
+ "learning_rate": 6.110877442254444e-05,
1107
+ "loss": 0.0379,
1108
+ "step": 1830
1109
+ },
1110
+ {
1111
+ "grad_norm": 0.8195701241493225,
1112
+ "learning_rate": 6.0705363343803946e-05,
1113
+ "loss": 0.0379,
1114
+ "step": 1840
1115
+ },
1116
+ {
1117
+ "grad_norm": 0.7879520654678345,
1118
+ "learning_rate": 6.030122056769934e-05,
1119
+ "loss": 0.0387,
1120
+ "step": 1850
1121
+ },
1122
+ {
1123
+ "grad_norm": 0.8573419451713562,
1124
+ "learning_rate": 5.989637371685257e-05,
1125
+ "loss": 0.0443,
1126
+ "step": 1860
1127
+ },
1128
+ {
1129
+ "grad_norm": 0.6557536721229553,
1130
+ "learning_rate": 5.949085046200808e-05,
1131
+ "loss": 0.0442,
1132
+ "step": 1870
1133
+ },
1134
+ {
1135
+ "grad_norm": 0.7351909279823303,
1136
+ "learning_rate": 5.908467852014169e-05,
1137
+ "loss": 0.0362,
1138
+ "step": 1880
1139
+ },
1140
+ {
1141
+ "grad_norm": 0.7618509531021118,
1142
+ "learning_rate": 5.867788565256607e-05,
1143
+ "loss": 0.0364,
1144
+ "step": 1890
1145
+ },
1146
+ {
1147
+ "grad_norm": 0.8200004696846008,
1148
+ "learning_rate": 5.827049966303335e-05,
1149
+ "loss": 0.0482,
1150
+ "step": 1900
1151
+ },
1152
+ {
1153
+ "grad_norm": 0.847327709197998,
1154
+ "learning_rate": 5.786254839583478e-05,
1155
+ "loss": 0.0421,
1156
+ "step": 1910
1157
+ },
1158
+ {
1159
+ "grad_norm": 0.8439093828201294,
1160
+ "learning_rate": 5.745405973389757e-05,
1161
+ "loss": 0.0394,
1162
+ "step": 1920
1163
+ },
1164
+ {
1165
+ "grad_norm": 0.7736167907714844,
1166
+ "learning_rate": 5.7045061596879134e-05,
1167
+ "loss": 0.0411,
1168
+ "step": 1930
1169
+ },
1170
+ {
1171
+ "grad_norm": 0.8771815896034241,
1172
+ "learning_rate": 5.6635581939258855e-05,
1173
+ "loss": 0.0397,
1174
+ "step": 1940
1175
+ },
1176
+ {
1177
+ "grad_norm": 0.7791712284088135,
1178
+ "learning_rate": 5.622564874842742e-05,
1179
+ "loss": 0.0404,
1180
+ "step": 1950
1181
+ },
1182
+ {
1183
+ "grad_norm": 0.8065997362136841,
1184
+ "learning_rate": 5.5815290042773836e-05,
1185
+ "loss": 0.0337,
1186
+ "step": 1960
1187
+ },
1188
+ {
1189
+ "grad_norm": 0.8823350667953491,
1190
+ "learning_rate": 5.540453386977058e-05,
1191
+ "loss": 0.041,
1192
+ "step": 1970
1193
+ },
1194
+ {
1195
+ "grad_norm": 0.7548691630363464,
1196
+ "learning_rate": 5.4993408304056425e-05,
1197
+ "loss": 0.0488,
1198
+ "step": 1980
1199
+ },
1200
+ {
1201
+ "grad_norm": 0.7795559167861938,
1202
+ "learning_rate": 5.458194144551768e-05,
1203
+ "loss": 0.0377,
1204
+ "step": 1990
1205
+ },
1206
+ {
1207
+ "grad_norm": 0.7108629941940308,
1208
+ "learning_rate": 5.417016141736756e-05,
1209
+ "loss": 0.034,
1210
+ "step": 2000
1211
+ }
1212
+ ],
1213
+ "logging_steps": 10,
1214
+ "max_steps": 4000,
1215
+ "num_input_tokens_seen": 0,
1216
+ "num_train_epochs": 9223372036854775807,
1217
+ "save_steps": 500,
1218
+ "stateful_callbacks": {
1219
+ "TrainerControl": {
1220
+ "args": {
1221
+ "should_epoch_stop": false,
1222
+ "should_evaluate": false,
1223
+ "should_log": false,
1224
+ "should_save": true,
1225
+ "should_training_stop": false
1226
+ },
1227
+ "attributes": {}
1228
+ }
1229
+ },
1230
+ "total_flos": 0.0,
1231
+ "train_batch_size": 32,
1232
+ "trial_name": null,
1233
+ "trial_params": null
1234
+ }
a14-unscrew-bulb/v2-b512/step-002000/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:569c50a90d82e42bacc2f68f87e2f4a3980093dc67dcea52a5497b16bad6b7cb
3
+ size 8017
a14-unscrew-bulb/v2-b512/step-002000/training_plan.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a14v2",
3
+ "frames": 185638,
4
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base",
5
+ "recipe": "full fine-tune (tune_llm, tune_visual, projector, diffusion, vlln), fixed steps, 2 nodes x 8 x 32, no accumulation",
6
+ "per_gpu_batch": 32,
7
+ "gradient_accumulation_steps": 1,
8
+ "num_nodes": 2,
9
+ "world_size": 16,
10
+ "global_batch": 512,
11
+ "max_steps": 4000,
12
+ "learning_rate": 0.0001,
13
+ "warmup_ratio": 0.05,
14
+ "lr_schedule": "linear warmup 5 %, cosine to 0",
15
+ "save_steps": 500,
16
+ "save_list": [
17
+ 2000,
18
+ 2500,
19
+ 3000,
20
+ 3500,
21
+ 4000
22
+ ],
23
+ "epochs_equivalent": 11.03,
24
+ "resume_note": "host-RAM growth (~0.5 GB per step on master-0) in the H2 video data pipeline: the run was stopped after checkpoint-2000 and checkpoint-3000 and resumed from them (--resume_from_checkpoint: weights, optimizer, scheduler, RNG; GR00T ignore_data_skip); recipe unchanged"
25
+ }
a14-unscrew-bulb/v2-b512/step-002000/wandb_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"project": "finetune-gr00t-n1d7", "run_id": "production"}
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/cache_cosmos_reason2.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Put the full gated nvidia/Cosmos-Reason2-2B (pinned 9ce19a1) into the shared HF hub cache $G/hf-home/hub.
2
+
3
+ GR00T N1.7 builds its backbone with Qwen3VLForConditionalGeneration.from_pretrained("nvidia/Cosmos-Reason2-2B") and
4
+ its processor with Qwen3VLProcessor.from_pretrained(same name) before loading the GR00T checkpoint, so every node needs
5
+ the whole repo resolvable by name. groot_node.sh runs with HF_HUB_OFFLINE=1, so nodes read this cache and never need
6
+ the token. refs/main is pinned to the verified revision.
7
+
8
+ Usage (token from stdin, kept in-process only):
9
+ $G/envs/groot-py312/bin/python cache_cosmos_reason2.py $G/hf-home/hub < /root/.hb-secrets/hf_token
10
+ """
11
+ import hashlib
12
+ import json
13
+ import os
14
+ import sys
15
+ import time
16
+
17
+ from huggingface_hub import HfApi, snapshot_download
18
+
19
+ REPO = "nvidia/Cosmos-Reason2-2B"
20
+ REV = "9ce19a195e423419c349abfc86fd07178b230561"
21
+ cache = os.path.abspath(sys.argv[1])
22
+ token = sys.stdin.readline().strip()
23
+ if not token.startswith("hf_"):
24
+ raise SystemExit("token format rejected")
25
+ api = HfApi(endpoint="https://huggingface.co", token=token)
26
+ info = api.model_info(REPO, revision=REV, files_metadata=True)
27
+ assert info.sha == REV, info.sha
28
+ t0 = time.monotonic()
29
+ for attempt in range(1, 21):
30
+ try:
31
+ path = snapshot_download(REPO, revision=REV, cache_dir=cache, token=token, max_workers=8)
32
+ break
33
+ except Exception as error: # transient stalls
34
+ if any(k in str(error).lower() for k in ("gated", "401", "403")):
35
+ raise
36
+ print(f"retry {attempt}: {type(error).__name__}: {str(error)[:200]}", flush=True)
37
+ time.sleep(10)
38
+ else:
39
+ raise SystemExit("download did not finish")
40
+ token = None
41
+ count, size = 0, 0
42
+ for s in info.siblings:
43
+ f = os.path.join(path, s.rfilename)
44
+ assert os.path.isfile(f), s.rfilename
45
+ if s.size is not None:
46
+ assert os.path.getsize(f) == s.size, s.rfilename
47
+ if s.lfs is not None:
48
+ h = hashlib.sha256()
49
+ with open(f, "rb") as stream:
50
+ for chunk in iter(lambda: stream.read(1 << 22), b""):
51
+ h.update(chunk)
52
+ assert h.hexdigest() == s.lfs.sha256, s.rfilename
53
+ count += 1
54
+ size += s.size or 0
55
+ refs = os.path.join(cache, "models--" + REPO.replace("/", "--"), "refs")
56
+ os.makedirs(refs, exist_ok=True)
57
+ with open(os.path.join(refs, "main"), "w") as f:
58
+ f.write(REV)
59
+ print("CACHE_OK", json.dumps({"repo": REPO, "revision": REV, "snapshot": path, "files": count, "bytes": size,
60
+ "seconds": round(time.monotonic() - t0, 1)}), flush=True)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/check_reason2_offline.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check that GR00T's Cosmos-Reason2-2B pieces load offline from $G/hf-home exactly as groot_node.sh will.
2
+
3
+ Usage: HF_HOME=$G/hf-home HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 $G/envs/groot-py312/bin/python check_reason2_offline.py
4
+ """
5
+ import json
6
+ import os
7
+ import sys
8
+ from pathlib import Path
9
+
10
+ import torch
11
+
12
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
13
+ import hb_offline_fix # noqa: E402,F401
14
+ from transformers import AutoConfig, Qwen3VLForConditionalGeneration, Qwen3VLProcessor
15
+
16
+ assert os.environ.get("HF_HUB_OFFLINE") == "1", "run with HF_HUB_OFFLINE=1"
17
+ name = "nvidia/Cosmos-Reason2-2B"
18
+ torch.set_num_threads(8)
19
+ proc = Qwen3VLProcessor.from_pretrained(name, trust_remote_code=True)
20
+ cfg = AutoConfig.from_pretrained(name, trust_remote_code=True)
21
+ ids = proc.tokenizer("pick up the trash and drop it in the bin")["input_ids"]
22
+ model = Qwen3VLForConditionalGeneration.from_pretrained(name, torch_dtype=torch.bfloat16)
23
+ n = sum(p.numel() for p in model.parameters())
24
+ print("REASON2_OFFLINE_OK", json.dumps({"processor": type(proc).__name__, "tokens": len(ids),
25
+ "model_type": cfg.model_type, "params_M": round(n / 1e6, 1),
26
+ "llm_layers": len(model.model.language_model.layers)}), flush=True)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/g1_feedback76_h16.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """G1 feedback76 / H16 GR00T modality config (copy of HumanWAM configs/humanwam/groot/feedback76_h16.py @ c76527b).
2
+
3
+ The matching modality.json is generated per dataset from the recorded joint names (prepare_groot_data.py);
4
+ HumanWAM's feedback76_modality.json has right_arm at observation.state[29:36] (left-hand fingers), which is wrong
5
+ for our G1 exports (right arm is [22:29]).
6
+ """
7
+ from gr00t.configs.data.embodiment_configs import register_modality_config
8
+ from gr00t.data.embodiment_tags import EmbodimentTag
9
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
10
+
11
+ feedback76_h16 = {
12
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["ego_view", "chest"]),
13
+ "state": ModalityConfig(
14
+ delta_indices=[0],
15
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "left_hand", "right_hand",
16
+ "projected_gravity"],
17
+ ),
18
+ "action": ModalityConfig(
19
+ delta_indices=list(range(16)),
20
+ modality_keys=["motion_token", "left_hand_joints", "right_hand_joints"],
21
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
22
+ format=ActionFormat.DEFAULT) for _ in range(3)],
23
+ ),
24
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
25
+ }
26
+
27
+ register_modality_config(feedback76_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/groot_chain.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Per-node GR00T relay: per-layout batch-size probe (largest that fits), then 20-epoch production per task.
2
+
3
+ Run on every node (nohup), node rank from the DLC env RANK: python groot_chain.py [task ...]
4
+ Default tasks: b6 a3 a7 a18 (basket postponed). A phase succeeds only if all 4 node launchers exit 0.
5
+
6
+ Batch (user rule: as large as memory allows, 2026-11-10): descending CANDIDATES, 4 steps each on all 32 GPUs; a
7
+ candidate fits when all launchers exit 0 AND the peak reserved memory of every rank (written by hb_launch.py to
8
+ logs/peak-<phase>-rank<r>.json) is <= MAX_FRAC of the GPU. Node 0 alone judges each probe and writes a verdict file
9
+ the other nodes wait for, so all nodes take the same decision. One decision per robot layout (g1 / h2).
10
+ Outputs: $G/groot-runs-bmax/ (the earlier per-GPU-64 runs stay in $G/groot-runs/).
11
+ """
12
+ import fcntl
13
+ import json
14
+ import math
15
+ import os
16
+ import shutil
17
+ import subprocess
18
+ import sys
19
+ import time
20
+ from pathlib import Path
21
+
22
+ H = Path(__file__).resolve().parent
23
+ G = H.parent
24
+ RUNS = G / os.environ.get("HB_GROOT_RUNS_NAME", "groot-runs-bmax")
25
+ NODE = int(os.environ["RANK"])
26
+ TASKS = sys.argv[1:] or ["b6", "a3", "a7", "a18"] # basket postponed (user plan 2026-11-10)
27
+ CANDIDATES = [320, 288, 256, 224, 192, 160, 128, 96, 64, 48, 32]
28
+ MAX_FRAC = 0.92
29
+ EPOCHS = 20
30
+ PORT_BASE = int(os.environ.get("HB_GROOT_PORT_BASE", "32500")) # probe PORT_BASE+i, production PORT_BASE+50+k
31
+ RUNS.mkdir(parents=True, exist_ok=True)
32
+ os.environ["HB_GROOT_RUNS"] = str(RUNS) # read by groot_node.sh
33
+ lock = (RUNS / f"chain-node-{NODE}.lock").open("a")
34
+ fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
35
+
36
+
37
+ def log(msg):
38
+ line = f"{time.strftime('%Y-%m-%d %H:%M:%S')} node{NODE} {msg}"
39
+ print(line, flush=True)
40
+ with (RUNS / f"chain-node-{NODE}.log").open("a") as f:
41
+ f.write(line + "\n")
42
+
43
+
44
+ def run_phase(task, phase, bs, steps, save, port):
45
+ logs = RUNS / task / "logs"
46
+ exits = [logs / f"{phase}-node-{r}.exit" for r in range(4)]
47
+ if all(p.is_file() for p in exits):
48
+ return all(p.read_text().strip() == "0" for p in exits)
49
+ subprocess.run(["bash", str(H / "groot_node.sh"), task, phase, str(bs), str(steps), str(save), str(port)])
50
+ deadline = time.monotonic() + 600
51
+ while not all(p.is_file() for p in exits):
52
+ if time.monotonic() > deadline:
53
+ log(f"{task}/{phase}: timed out waiting for peer exit files")
54
+ return False
55
+ time.sleep(5)
56
+ ok = all(p.read_text().strip() == "0" for p in exits)
57
+ log(f"{task}/{phase}: bs={bs} steps={steps} -> {'OK' if ok else 'FAIL'}")
58
+ return ok
59
+
60
+
61
+ def peak_frac(task, phase):
62
+ """Max over ranks of peak reserved / total; None when no rank reported."""
63
+ fracs = []
64
+ for f in (RUNS / task / "logs").glob(f"peak-{phase}-rank*.json"):
65
+ try:
66
+ d = json.loads(f.read_text())
67
+ fracs.append(d["peak_reserved_gb"] / d["total_gb"])
68
+ except (ValueError, KeyError, OSError):
69
+ continue
70
+ return (max(fracs), len(fracs)) if fracs else (None, 0)
71
+
72
+
73
+ def probe_verdict(task, cand, port):
74
+ phase = f"probe-bs{cand}"
75
+ verdict = RUNS / task / "logs" / f"{phase}.verdict.json"
76
+ ok = run_phase(task, phase, cand, 4, 1000000, port)
77
+ if NODE == 0 and not verdict.is_file():
78
+ time.sleep(10) # let peak files from the other nodes become visible on NAS
79
+ frac, n = peak_frac(task, phase) if ok else (None, 0)
80
+ fits = bool(ok and frac is not None and frac <= MAX_FRAC)
81
+ verdict.write_text(json.dumps({"per_gpu_batch": cand, "ran_ok": ok, "peak_frac": frac, "ranks_reported": n,
82
+ "max_frac": MAX_FRAC, "fits": fits}) + "\n")
83
+ deadline = time.monotonic() + 600
84
+ while not verdict.is_file():
85
+ if time.monotonic() > deadline:
86
+ log(f"{task}/{phase}: no verdict from node 0")
87
+ return False
88
+ time.sleep(5)
89
+ v = json.loads(verdict.read_text())
90
+ log(f"{task}/{phase}: verdict {v}")
91
+ return v["fits"]
92
+
93
+
94
+ def prune_resume_state(out):
95
+ """Keep weights of every checkpoint but resume state (DeepSpeed optimizer shards, scheduler, RNG) only in the latest."""
96
+ ckpts = sorted((p for p in out.glob("checkpoint-*") if p.name.split("-")[-1].isdigit()),
97
+ key=lambda p: int(p.name.split("-")[-1]))
98
+ for c in ckpts[:-1]:
99
+ for p in list(c.glob("global_step*")) + list(c.glob("optimizer*.pt")) + list(c.glob("scheduler.pt")) \
100
+ + list(c.glob("rng_state*.pth")) + list(c.glob("latest")) + list(c.glob("zero_to_fp32.py")):
101
+ if p.is_dir():
102
+ shutil.rmtree(p, ignore_errors=True)
103
+ else:
104
+ p.unlink(missing_ok=True)
105
+ log(f"{c.name}: resume state dropped (weights kept)")
106
+
107
+
108
+ def batch_for(task):
109
+ layout = "h2" if task in ("a17", "a18") else "g1"
110
+ decision = RUNS / f"batch-decision-{layout}.json"
111
+ if decision.is_file():
112
+ return json.loads(decision.read_text())["per_gpu_batch"]
113
+ for i, cand in enumerate(CANDIDATES):
114
+ if probe_verdict(task, cand, PORT_BASE + i + (20 if layout == "h2" else 0)):
115
+ if NODE == 0:
116
+ decision.write_text(json.dumps({"per_gpu_batch": cand, "global_batch": cand * 32, "probe_task": task,
117
+ "layout": layout, "candidates": CANDIDATES,
118
+ "max_frac": MAX_FRAC}) + "\n")
119
+ return cand
120
+ time.sleep(20)
121
+ log(f"{layout}: no batch size fit; stopping")
122
+ sys.exit(1)
123
+
124
+
125
+ for k, task in enumerate(TASKS):
126
+ bs = batch_for(task)
127
+ log(f"{task}: per-GPU batch {bs} (global {bs * 32})")
128
+ receipt = json.loads((G / "groot-data" / task / "hb-dataset-receipt.json").read_text())
129
+ steps = math.ceil(EPOCHS * receipt["frames"] / (bs * 32))
130
+ save = math.ceil(steps / 2)
131
+ if not run_phase(task, "production", bs, steps, save, PORT_BASE + 50 + k):
132
+ log(f"{task}: production failed; stopping chain")
133
+ sys.exit(1)
134
+ if NODE == 0:
135
+ prune_resume_state(RUNS / task / "production")
136
+ (RUNS / task / "production-complete.json").write_text(json.dumps(
137
+ {"task": task, "frames": receipt["frames"], "epochs": EPOCHS, "per_gpu_batch": bs,
138
+ "global_batch": bs * 32, "max_steps": steps, "save_steps": save}) + "\n")
139
+ log("all tasks done")
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/h2_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A18 H2 (Sharpa hands) GR00T modality config, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/h2_body31_hand44_planner7_h16.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 v2 H2 (Sharpa hands) GR00T modality config (H2 training guide hb_training/docs/h2/a14_training_guide.md 4.1):
2
+ cameras head + chest, state78 = 75 joints + projected gravity, action82 = action.body_target (31) + action.hands (44)
3
+ + action.planner (7, the SONIC planner command), horizon 16, absolute targets. Identical to h2_body31_hand44_h16.py
4
+ except the added `planner` action modality."""
5
+ from gr00t.configs.data.embodiment_configs import register_modality_config
6
+ from gr00t.data.embodiment_tags import EmbodimentTag
7
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
8
+
9
+ h2_body31_hand44_planner7_h16 = {
10
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest"]),
11
+ "state": ModalityConfig(
12
+ delta_indices=[0],
13
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
14
+ "projected_gravity"],
15
+ ),
16
+ "action": ModalityConfig(
17
+ delta_indices=list(range(16)),
18
+ modality_keys=["body_target", "left_hand_target", "right_hand_target", "planner"],
19
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
20
+ format=ActionFormat.DEFAULT) for _ in range(4)],
21
+ ),
22
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
23
+ }
24
+
25
+ register_modality_config(h2_body31_hand44_planner7_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/h2w3_body31_hand44_h16.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A14 H2 (Sharpa hands) GR00T modality config with THREE cameras head + chest + wrist (user 2026-11-12; the third-person camera is never a model input). Otherwise identical to h2_body31_hand44_h16.py:, same protocol as the A18 Delta-flow run
2
+ (schema a18-h2-sharpa-body31-hand44-state78-chunk16-v1): cameras head+chest, state = 75 joints + projected
3
+ gravity, action = action.body_target (31) + action.hands (44), horizon 16, absolute targets."""
4
+ from gr00t.configs.data.embodiment_configs import register_modality_config
5
+ from gr00t.data.embodiment_tags import EmbodimentTag
6
+ from gr00t.data.types import ActionConfig, ActionFormat, ActionRepresentation, ActionType, ModalityConfig
7
+
8
+ h2w3_body31_hand44_h16 = {
9
+ "video": ModalityConfig(delta_indices=[0], modality_keys=["head", "chest", "wrist"]),
10
+ "state": ModalityConfig(
11
+ delta_indices=[0],
12
+ modality_keys=["left_leg", "right_leg", "waist", "left_arm", "right_arm", "head", "left_hand", "right_hand",
13
+ "projected_gravity"],
14
+ ),
15
+ "action": ModalityConfig(
16
+ delta_indices=list(range(16)),
17
+ modality_keys=["body_target", "left_hand_target", "right_hand_target"],
18
+ action_configs=[ActionConfig(rep=ActionRepresentation.ABSOLUTE, type=ActionType.NON_EEF,
19
+ format=ActionFormat.DEFAULT) for _ in range(3)],
20
+ ),
21
+ "language": ModalityConfig(delta_indices=[0], modality_keys=["annotation.human.task_description"]),
22
+ }
23
+
24
+ register_modality_config(h2w3_body31_hand44_h16, embodiment_tag=EmbodimentTag.NEW_EMBODIMENT)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/hb_aug_v2.py ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Humanoid Benchmark data augmentation v2 (project CLAUDE.md), model-agnostic torch implementation.
2
+
3
+ augment(x, rng, obs_mask=None) takes x: float [T, V, 3, H, W] in [0, 1] (T frames, V cameras) and returns the
4
+ augmented tensor. Each call is one training sample: one of the 15 effects is chosen uniformly and its magnitude
5
+ is scaled by 1.5 around identity; the rest stay at x1. Colour is shared across cameras and time; geometry, light
6
+ patches and white balance are drawn per camera and fixed over time; sensor/optics effects (noise, gaussian blur,
7
+ JPEG, motion blur, sharpen, defocus) apply only to observation frames (obs_mask True).
8
+ Pixel-size parameters are defined at 224 px and scaled by H / 224.
9
+ perturb_joints(q, rng) applies the proprio spec: per-joint bias +-2 deg plus N(0, 0.25 deg) clipped at 3 sigma.
10
+ """
11
+ import io
12
+ import math
13
+
14
+ import numpy as np
15
+ import torch
16
+ import torch.nn.functional as F
17
+ from PIL import Image
18
+ from torchvision.transforms import InterpolationMode
19
+ from torchvision.transforms import functional as TF
20
+
21
+ EFFECTS = ("exposure", "contrast", "saturation", "hue", "gamma", "temperature", "rotation_crop", "noise",
22
+ "gaussian_blur", "jpeg", "motion_blur", "sharpen", "defocus", "light_shadow", "wb_channel")
23
+ CENTER = {"exposure": 0.0, "contrast": 1.0, "saturation": 1.0, "hue": 0.0, "gamma": 1.0, "temperature": 0.0}
24
+ RANGE = {"exposure": (-0.22, 0.22), "contrast": (0.88, 1.12), "saturation": (0.76, 1.24), "hue": (-0.015, 0.015),
25
+ "gamma": (0.89, 1.11), "temperature": (-0.066, 0.066)}
26
+ VERSION = "hb-aug-v2-20261109"
27
+
28
+
29
+ def sample(rng, views, height):
30
+ boost = EFFECTS[int(rng.integers(len(EFFECTS)))]
31
+
32
+ def k(name):
33
+ return 1.5 if boost == name else 1.0
34
+
35
+ px = height / 224.0
36
+ p = {"boost": boost}
37
+ for name, (a, b) in RANGE.items():
38
+ c = CENTER[name]
39
+ p[name] = float(rng.uniform(c + k(name) * (a - c), c + k(name) * (b - c)))
40
+ p["geometry"] = []
41
+ for _ in range(views):
42
+ angle = float(rng.uniform(-1.5 * k("rotation_crop"), 1.5 * k("rotation_crop")))
43
+ p["geometry"].append((angle, 1.0 + abs(angle) * 0.015 + float(rng.uniform(0, 0.02 * k("rotation_crop")))))
44
+ p["noise"] = 4.0 / 255.0 * k("noise") if rng.random() < 0.5 else 0.0
45
+ p["gaussian_blur"] = float(rng.uniform(0.25, 0.65)) * k("gaussian_blur") if rng.random() < 0.15 else 0.0
46
+ if rng.random() < 0.35:
47
+ lo = int(round(100 - k("jpeg") * 30))
48
+ hi = int(round(100 - k("jpeg") * 5))
49
+ p["jpeg"] = int(rng.integers(lo, hi + 1))
50
+ else:
51
+ p["jpeg"] = None
52
+ p["motion_blur"] = ((float(rng.uniform(3, 7)) * k("motion_blur") * px, float(rng.uniform(0, 180)))
53
+ if rng.random() < 0.15 else None)
54
+ p["sharpen"] = float(rng.uniform(0.2, 0.6)) * k("sharpen") if rng.random() < 0.15 else 0.0
55
+ p["defocus"] = float(rng.uniform(1, 2)) * k("defocus") * px if rng.random() < 0.10 else 0.0
56
+ if rng.random() < 0.30:
57
+ g = 0.30 * k("light_shadow")
58
+ p["light_shadow"] = [[(float(rng.uniform(0.1, 0.9)), float(rng.uniform(0.1, 0.9)),
59
+ float(rng.uniform(0.15, 0.35)), float(rng.uniform(1 - g, 1 + g)))
60
+ for _ in range(int(rng.integers(1, 3)))] for _ in range(views)]
61
+ else:
62
+ p["light_shadow"] = None
63
+ if rng.random() < 0.30:
64
+ g = 0.10 * k("wb_channel")
65
+ gains = []
66
+ for _ in range(views):
67
+ row = [1.0, 1.0, 1.0]
68
+ row[int(rng.integers(0, 3))] = float(rng.uniform(1 - g, 1 + g))
69
+ gains.append(row)
70
+ p["wb_channel"] = gains
71
+ else:
72
+ p["wb_channel"] = None
73
+ return p
74
+
75
+
76
+ def _conv(x, kernel):
77
+ n, c, h, w = x.shape
78
+ k = kernel.to(x)[None, None].repeat(c, 1, 1, 1)
79
+ pad = kernel.shape[-1] // 2
80
+ return F.conv2d(F.pad(x, (pad, pad, pad, pad), mode="reflect"), k, groups=c)
81
+
82
+
83
+ def motion_kernel(length, angle_deg):
84
+ size = max(3, int(round(length)) | 1)
85
+ k = torch.zeros(size, size)
86
+ c = size // 2
87
+ th = math.radians(angle_deg)
88
+ for s in np.linspace(-(length - 1) / 2, (length - 1) / 2, 8 * size):
89
+ k[int(round(c - s * math.sin(th))), int(round(c + s * math.cos(th)))] = 1.0
90
+ return k / k.sum()
91
+
92
+
93
+ def disk_kernel(r):
94
+ size = 2 * int(math.ceil(r)) + 1
95
+ c = size // 2
96
+ yy, xx = torch.meshgrid(torch.arange(size), torch.arange(size), indexing="ij")
97
+ k = (((yy - c) ** 2 + (xx - c) ** 2) <= (r + 0.5) ** 2).float()
98
+ return k / k.sum()
99
+
100
+
101
+ def apply(x, p, obs_mask=None, noise_seed=0):
102
+ t, v, c, h, w = x.shape
103
+ x = x.clone().clamp(0, 1)
104
+ for vi, (angle, scale) in enumerate(p["geometry"]):
105
+ x[:, vi] = TF.affine(x[:, vi], angle=angle, translate=[0, 0], scale=scale, shear=[0, 0],
106
+ interpolation=InterpolationMode.BILINEAR)
107
+ if p["light_shadow"] is not None:
108
+ yy, xx = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w), indexing="ij")
109
+ for vi, blobs in enumerate(p["light_shadow"]):
110
+ m = torch.ones(h, w)
111
+ for cx, cy, s, g in blobs:
112
+ m = m + (g - 1.0) * torch.exp(-((xx - cx) ** 2 + (yy - cy) ** 2) / (2 * s * s))
113
+ x[:, vi] = (x[:, vi] * m.clamp(min=0).to(x)).clamp(0, 1)
114
+ if p["wb_channel"] is not None:
115
+ for vi, g in enumerate(p["wb_channel"]):
116
+ x[:, vi] = (x[:, vi] * torch.tensor(g, dtype=x.dtype).view(1, 3, 1, 1)).clamp(0, 1)
117
+ y = x.reshape(t * v, c, h, w)
118
+ y = (y * float(2.0 ** p["exposure"])).clamp(0, 1)
119
+ d = p["temperature"]
120
+ y = (y * y.new_tensor([1.0 + d, 1.0, 1.0 - d]).view(1, 3, 1, 1)).clamp(0, 1)
121
+ y = ((y - 0.5) * p["contrast"] + 0.5).clamp(0, 1)
122
+ y = TF.adjust_saturation(y, p["saturation"])
123
+ y = TF.adjust_hue(y, p["hue"])
124
+ y = y.clamp(0, 1).pow(p["gamma"])
125
+ x = y.reshape(t, v, c, h, w)
126
+ mask = torch.ones(t, dtype=torch.bool) if obs_mask is None else torch.as_tensor(obs_mask, dtype=torch.bool)
127
+ if bool(mask.any()):
128
+ o = x[mask].reshape(-1, c, h, w)
129
+ if p["gaussian_blur"] > 0:
130
+ o = TF.gaussian_blur(o, [3, 3], [p["gaussian_blur"]] * 2)
131
+ if p["motion_blur"] is not None:
132
+ o = _conv(o, motion_kernel(*p["motion_blur"]))
133
+ if p["defocus"] > 0:
134
+ o = _conv(o, disk_kernel(p["defocus"]))
135
+ if p["sharpen"] > 0:
136
+ o = o + p["sharpen"] * (o - TF.gaussian_blur(o, [5, 5], [1.0, 1.0]))
137
+ o = o.clamp(0, 1)
138
+ if p["noise"] > 0:
139
+ gen = torch.Generator().manual_seed(int(noise_seed))
140
+ o = (o + torch.randn(o.shape, generator=gen, dtype=o.dtype) * p["noise"]).clamp(0, 1)
141
+ if p["jpeg"] is not None:
142
+ for i in range(o.shape[0]):
143
+ buf = io.BytesIO()
144
+ TF.to_pil_image(o[i]).save(buf, format="JPEG", quality=int(p["jpeg"]))
145
+ buf.seek(0)
146
+ with Image.open(buf) as im:
147
+ o[i] = TF.pil_to_tensor(im.convert("RGB")).to(o.dtype).div(255.0)
148
+ x[mask] = o.reshape(-1, v, c, h, w)
149
+ return x.clamp(0, 1)
150
+
151
+
152
+ def augment(x, rng, obs_mask=None):
153
+ p = sample(rng, x.shape[1], x.shape[-2])
154
+ return apply(x, p, obs_mask, noise_seed=int(rng.integers(0, 2 ** 62))), p
155
+
156
+
157
+ def perturb_joints(q, rng):
158
+ q = np.asarray(q, dtype=np.float64)
159
+ d = q.shape[-1]
160
+ bias = rng.uniform(-math.radians(2.0), math.radians(2.0), size=d)
161
+ sigma = math.radians(0.25)
162
+ noise = np.clip(rng.normal(0.0, sigma, size=q.shape), -3 * sigma, 3 * sigma)
163
+ return (q + bias + noise).astype(np.float32)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/hb_launch.py ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GR00T N1.7 fine-tuning entry with Humanoid Benchmark augmentation v2.
2
+
3
+ Patches Gr00tN1d7Processor in training mode only:
4
+ - images: native random crop / colour jitter / rotation are bypassed (the deterministic eval resize+center-crop is
5
+ used), then hb_aug_v2 is applied jointly to all cameras of the sample;
6
+ - proprio: the body-joint state groups (HB_BODY_STATE_KEYS) get the v2 bias + noise before normalization.
7
+ Then runs the pinned upstream gr00t/experiment/launch_finetune.py unchanged.
8
+ """
9
+ import json
10
+ import os
11
+ import runpy
12
+ import sys
13
+ from pathlib import Path
14
+
15
+ import numpy as np
16
+ import torch
17
+
18
+ HERE = Path(__file__).resolve().parent
19
+ sys.path.insert(0, str(HERE))
20
+ import hb_offline_fix # noqa: E402,F401 (tokenizer load under HF_HUB_OFFLINE=1, see module docstring)
21
+ import hb_aug_v2 as A # noqa: E402
22
+ from gr00t.model.gr00t_n1d7 import processing_gr00t_n1d7 as P # noqa: E402
23
+
24
+ Proc = P.Gr00tN1d7Processor
25
+ BODY_KEYS = [k for k in os.environ.get("HB_BODY_STATE_KEYS", "left_leg,right_leg,waist,left_arm,right_arm").split(",") if k]
26
+ LOG = os.environ.get("HB_AUG_LOG")
27
+ _state = {"rng": None, "count": 0}
28
+
29
+
30
+ def _rng():
31
+ if _state["rng"] is None:
32
+ _state["rng"] = np.random.default_rng([os.getpid(), int.from_bytes(os.urandom(8), "little")])
33
+ return _state["rng"]
34
+
35
+
36
+ _orig_get, _orig_apply, _orig_call = Proc._get_vlm_inputs, Proc._apply_vlm_processing, Proc.__call__
37
+
38
+
39
+ def _get_vlm_inputs(self, image_keys, images, masks, image_transform, language):
40
+ if getattr(self, "training", False):
41
+ self._hb_views = len(image_keys)
42
+ try:
43
+ return _orig_get(self, image_keys, images, masks, self.eval_image_transform, language)
44
+ finally:
45
+ self._hb_views = None
46
+ return _orig_get(self, image_keys, images, masks, image_transform, language)
47
+
48
+
49
+ def _apply_vlm_processing(self, images, language):
50
+ views = getattr(self, "_hb_views", None)
51
+ if views and isinstance(images, torch.Tensor):
52
+ shape = images.shape
53
+ x = images.reshape(-1, views, *shape[1:]).float().div(255.0)
54
+ x, p = A.augment(x, _rng())
55
+ images = (x * 255.0 + 0.5).clamp(0, 255).to(torch.uint8).reshape(shape)
56
+ _state["count"] += 1
57
+ if LOG and _state["count"] in (1, 50):
58
+ with open(f"{LOG}.{os.getpid()}", "a") as f:
59
+ f.write(json.dumps({"count": _state["count"], "views": views, "hw": list(shape[-2:]),
60
+ "boost": p["boost"], "version": A.VERSION}) + "\n")
61
+ return _orig_apply(self, images, language)
62
+
63
+
64
+ def _call(self, messages):
65
+ if getattr(self, "training", False) and BODY_KEYS:
66
+ states = messages[0]["content"].states
67
+ for key in BODY_KEYS:
68
+ if key in states:
69
+ states[key] = A.perturb_joints(states[key], _rng())
70
+ return _orig_call(self, messages)
71
+
72
+
73
+ Proc._get_vlm_inputs = _get_vlm_inputs
74
+ Proc._apply_vlm_processing = _apply_vlm_processing
75
+ Proc.__call__ = _call
76
+
77
+ if os.environ.get("HB_PATCH_ONLY") != "1":
78
+ entry = Path(os.environ["GROOT_SRC"]) / "gr00t/experiment/launch_finetune.py"
79
+ sys.argv = [str(entry)] + sys.argv[1:]
80
+ try:
81
+ runpy.run_path(str(entry), run_name="__main__")
82
+ finally: # batch probe: report this rank's peak reserved memory (groot_chain.py compares it with MAX_FRAC)
83
+ _peak = os.environ.get("HB_PEAK_FILE")
84
+ if _peak and torch.cuda.is_available():
85
+ try:
86
+ _dev = torch.cuda.current_device()
87
+ with open(f"{_peak}-rank{os.environ.get('RANK', '0')}.json", "w") as _f:
88
+ json.dump({"rank": int(os.environ.get("RANK", "0")),
89
+ "peak_reserved_gb": torch.cuda.max_memory_reserved(_dev) / 1e9,
90
+ "total_gb": torch.cuda.get_device_properties(_dev).total_memory / 1e9}, _f)
91
+ except Exception: # noqa: BLE001 - never mask the training result
92
+ pass
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/hb_offline_fix.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Make tokenizer loading work with HF_HUB_OFFLINE=1 on transformers 4.57.x.
2
+
3
+ PreTrainedTokenizerBase._patch_mistral_regex calls huggingface_hub.model_info() for every hub-named tokenizer with a
4
+ vocab > 100k (to detect Mistral base models). With HF_HUB_OFFLINE=1 that raises OfflineModeIsEnabled even though all
5
+ files are in the cache. GR00T's Cosmos-Reason2-2B / Qwen3-VL tokenizer is not a Mistral tokenizer, so when offline the
6
+ check is skipped and the tokenizer is returned unchanged. Online behaviour is untouched.
7
+ """
8
+ import os
9
+
10
+ from transformers.tokenization_utils_base import PreTrainedTokenizerBase as _PTB
11
+
12
+ if not getattr(_PTB, "_hb_offline_fix", False):
13
+ _orig = _PTB._patch_mistral_regex.__func__
14
+
15
+ def _patch_mistral_regex(cls, tokenizer, *args, **kwargs):
16
+ try:
17
+ return _orig(cls, tokenizer, *args, **kwargs)
18
+ except Exception as error:
19
+ if os.environ.get("HF_HUB_OFFLINE") == "1" or type(error).__name__ == "OfflineModeIsEnabled":
20
+ return tokenizer
21
+ raise
22
+
23
+ _PTB._patch_mistral_regex = classmethod(_patch_mistral_regex)
24
+ _PTB._hb_offline_fix = True
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/prepare_groot_data.py ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Build GR00T dataset views (symlinked episodes + rewritten meta + stats) for the G1 tasks.
2
+
3
+ Usage: <groot python> prepare_groot_data.py <task> [<task> ...] (tasks: b6 basket a3 a7)
4
+ B6 always uses the fixed first 100 episodes (0-99) of b6-v5t7-ready-start-20261109.
5
+ """
6
+ import hashlib
7
+ import json
8
+ import os
9
+ import subprocess
10
+ import sys
11
+ from pathlib import Path
12
+
13
+ H = Path(__file__).resolve().parent
14
+ G = H.parent
15
+ P = G.parent
16
+ OUT = G / "groot-data"
17
+ TASKS = {
18
+ "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100),
19
+ "basket": (P / "basket-current-dynamics-100-v21-b0ff522/data", None), # HF basket-transfer-expert-v1 @current-dynamics-100-v21 b0ff522 (old ed77719 retired)
20
+ "basketv2": (P / "basket-trimv2-c4df4d4/data", None), # HF a20-basket-transfer-pre-reach-trim-expert-v2 @c4df4d40 (user 2026-11-14)
21
+ "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100), # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
22
+ "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100), # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
23
+ "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None), # HF door-pull-expert-v1 @cf3ee9b (A6)
24
+ "b7green": (P / "b7-green-clothes-hamper-v1-ada288c/data", None), # HF b7-green-clothes-hamper-expert-v1 @ada288c, 100 episodes (user 2026-11-13)
25
+ "a7v2": (P / "a7-cabinet-close-v2-aecf070/data", None), # HF a7-cabinet-close-expert-v2 @aecf070, 100 episodes (user 2026-11-14)
26
+ "b8": (P / "b8-table-wipe-v1-3b2e618/data", None), # HF b8-table-wipe-expert-v1 @3b2e618, 100 episodes (user 2026-11-14)
27
+ "b1": (P / "b1-riceball-microwave-v1-36da298/data", None), # HF b1-riceball-microwave-expert-v1 @36da298, all 115 episodes (user 2026-11-14)
28
+ "b7greenv2": (P / "b7-green-clothes-hamper-v2-c12757e/data", None), # HF b7-green-clothes-hamper @c12757e (v2 at the repo root), all 120 episodes (user 2026-11-14)
29
+ "a10v2": (P / "a10-trolley-push-v2-d3abb4c/data", None), # HF a10-trolley-push-expert-v2 @d3abb4c, all 150 episodes (user 2026-11-13)
30
+ "a1": (P / "a1-ground-object-pickup-v1-f47f07e/data", None), # HF a1-ground-object-pickup-expert-v1 @f47f07e (A1)
31
+ "c6t": (P / "c6-test8-06c231d/data", None), # c6 8-episode adaptation test only
32
+ "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None), # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
33
+ "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
34
+ "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
35
+ "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
36
+ "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # HF a17-unscrew-bottle-expert-v1 @407e6e2
37
+ "a18c2": (P / "a18-h2-dataset-audit-8e956a10/dataset", None), # A18, 2 cameras head + chest (user 2026-11-14 18:2x)
38
+ "a17c2": (P / "a17-unscrew-bottle-v1-407e6e2/data", None), # A17, 2 cameras head + chest (user 2026-11-14 18:2x)
39
+ "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None), # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
40
+ "a14v2": (P / "a14-unscrew-bulb-v2-8411a6d/data", None), # HF a14-unscrew-bulb-expert-v2 @8411a6da: + action.planner (H2 guide 2026-11-14)
41
+ }
42
+ CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py",
43
+ "a18c2": "h2_body31_hand44_h16.py", "a17c2": "h2_body31_hand44_h16.py",
44
+ "a14v2": "h2_body31_hand44_planner7_h16.py"} # 2026-11-14 H2 guide: action82 incl. the planner command
45
+ # model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
46
+ CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist"),
47
+ "a18c2": ("head", "chest"), "a17c2": ("head", "chest"), "a14v2": ("head", "chest")} # 2026-11-14: 2-camera re-train (user)
48
+ PLANNER = {"a14v2"} # tasks whose action adds action.planner (7) after body_target + hands (H2 guide a14 2.2)
49
+ H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
50
+ "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}
51
+
52
+
53
+ def h2_modality(info, cams=("head", "chest")):
54
+ f = info["features"]
55
+ names = f["observation.state"]["names"]
56
+ assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
57
+ assert f["observation.projected_gravity"]["shape"] == [3]
58
+ state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
59
+ state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
60
+ return {"state": state,
61
+ "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
62
+ "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
63
+ "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
64
+ "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
65
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
66
+ BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
67
+ "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
68
+ "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
69
+ "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
70
+ "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
71
+ "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
72
+ "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
73
+ GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}
74
+
75
+
76
+ def jl(path):
77
+ return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]
78
+
79
+
80
+ def wl(path, rows):
81
+ Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))
82
+
83
+
84
+ def build(task):
85
+ src, limit = TASKS[task]
86
+ info = json.loads((src / "meta/info.json").read_text())
87
+ names = info["features"]["observation.state"]["names"]
88
+ if task in CONFIG:
89
+ idx = list(range(31))
90
+ else:
91
+ idx = [names.index(n) for n in BODY]
92
+ for g, (a, b) in GROUPS.items():
93
+ assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
94
+ for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
95
+ ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
96
+ ("observation.images.ego_view", None), ("observation.images.chest", None)):
97
+ assert key in info["features"], (task, key)
98
+ if dim:
99
+ assert info["features"][key]["shape"] == [dim], (task, key)
100
+ episodes = jl(src / "meta/episodes.jsonl")
101
+ stats = jl(src / "meta/episodes_stats.jsonl")
102
+ n = len(episodes) if limit is None else limit
103
+ episodes, stats = episodes[:n], stats[:n]
104
+ assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
105
+ frames = sum(int(e["length"]) for e in episodes)
106
+ dst = OUT / task
107
+ if dst.exists():
108
+ raise SystemExit(f"{dst} exists")
109
+ (dst / "meta").mkdir(parents=True)
110
+ vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
111
+ for e in episodes:
112
+ i = e["episode_index"]
113
+ chunk = i // int(info["chunks_size"])
114
+ rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
115
+ (dst / rel).parent.mkdir(parents=True, exist_ok=True)
116
+ os.symlink(src / rel, dst / rel)
117
+ for k in vkeys:
118
+ relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
119
+ (dst / relv).parent.mkdir(parents=True, exist_ok=True)
120
+ os.symlink(src / relv, dst / relv)
121
+ if task in CONFIG:
122
+ # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
123
+ # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
124
+ keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
125
+ "timestamp", "frame_index", "episode_index", "index", "task_index"}
126
+ keep |= {f"observation.images.{c}" for c in CAMS[task]}
127
+ keep |= {"action.planner"} if task in PLANNER else set()
128
+ info["features"] = {k: v for k, v in info["features"].items() if k in keep}
129
+ vkeys = [k for k in vkeys if k in keep]
130
+ stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
131
+ for s in stats]
132
+ info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
133
+ total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
134
+ (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
135
+ wl(dst / "meta/episodes.jsonl", episodes)
136
+ wl(dst / "meta/episodes_stats.jsonl", stats)
137
+ (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
138
+ if task in CONFIG:
139
+ modality = h2_modality(info, CAMS[task])
140
+ if task in PLANNER:
141
+ assert info["features"]["action.planner"]["shape"] == [7]
142
+ modality["action"]["planner"] = {"original_key": "action.planner", "start": 0, "end": 7}
143
+ else:
144
+ state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
145
+ for g, (a, b) in GROUPS.items()}
146
+ state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
147
+ right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
148
+ projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
149
+ modality = {"state": state,
150
+ "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
151
+ "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
152
+ "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
153
+ "video": {"ego_view": {"original_key": "observation.images.ego_view"},
154
+ "chest": {"original_key": "observation.images.chest"}},
155
+ "annotation": {"human.task_description": {"original_key": "task_index"}}}
156
+ (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
157
+ py = sys.executable
158
+ src_g = G / "HumanWAM/third_party/Isaac-GR00T"
159
+ env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
160
+ done = None
161
+ for tag in ("NEW_EMBODIMENT", "new_embodiment"):
162
+ r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
163
+ "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
164
+ capture_output=True, text=True)
165
+ if r.returncode == 0:
166
+ done = tag
167
+ break
168
+ last = r.stdout[-1500:] + r.stderr[-1500:]
169
+ if done is None:
170
+ raise SystemExit(f"stats failed for {task}:\n{last}")
171
+ assert (dst / "meta/stats.json").is_file()
172
+ receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
173
+ state_dim_raw=len(names), body_index=idx, stats_tag=done,
174
+ modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
175
+ stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
176
+ (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
177
+ print("DATASET_OK", json.dumps(receipt)[:400], flush=True)
178
+
179
+
180
+ for t in sys.argv[1:]:
181
+ build(t)
a14-unscrew-bulb/v2-b512/step-004000/code/hb_groot/verify_groot_cpu.py ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """CPU check of the GR00T data path for one task view: loader shapes, v2 augmentation, processor patch import."""
2
+ import json
3
+ import os
4
+ import sys
5
+ from pathlib import Path
6
+
7
+ import numpy as np
8
+ import torch
9
+
10
+ H = Path(__file__).resolve().parent
11
+ sys.path.insert(0, str(H))
12
+ os.environ["HB_PATCH_ONLY"] = "1"
13
+ import hb_launch # noqa: E402,F401 (applies the processor patch)
14
+ import hb_aug_v2 as A # noqa: E402
15
+ from gr00t.data.dataset.lerobot_episode_loader import LeRobotEpisodeLoader # noqa: E402
16
+ from gr00t.model.gr00t_n1d7.processing_gr00t_n1d7 import Gr00tN1d7Processor # noqa: E402
17
+
18
+ view = Path(sys.argv[1])
19
+ if view.name == "a14v2": # 2026-11-14 H2 guide: head + chest, action82 = body_target + hands + planner
20
+ import h2_body31_hand44_planner7_h16 as C # noqa: E402
21
+ cfg = C.h2_body31_hand44_planner7_h16
22
+ elif view.name in ("a18c2", "a17c2"): # 2026-11-14: H2 2-camera re-train (head + chest)
23
+ import h2_body31_hand44_h16 as C # noqa: E402
24
+ cfg = next(v for k, v in vars(C).items() if isinstance(v, dict) and "video" in v)
25
+ elif view.name in ("a14", "a18", "a17"): # every H2 task: 3 cameras head + chest + wrist (user 2026-11-12)
26
+ import h2w3_body31_hand44_h16 as C # noqa: E402
27
+ cfg = C.h2w3_body31_hand44_h16
28
+ else:
29
+ import g1_feedback76_h16 as C # noqa: E402
30
+ cfg = C.feedback76_h16
31
+ loader = LeRobotEpisodeLoader(view, cfg)
32
+ df = loader[0]
33
+ cols = list(df.columns)
34
+ row = df.iloc[20]
35
+ out = {"episodes": len(loader), "columns": cols[:40], "rows_ep0": len(df)}
36
+ state_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("state.")}
37
+ action_dims = {c: int(np.asarray(row[c]).size) for c in cols if c.startswith("action.")}
38
+ out.update(state_dims=state_dims, action_dims=action_dims, state_total=sum(state_dims.values()),
39
+ action_total=sum(action_dims.values()))
40
+ vids = [c for c in cols if c.startswith("video.")]
41
+ imgs = [np.asarray(row[c]) for c in vids]
42
+ out["video"] = {c: list(i.shape) for c, i in zip(vids, imgs)}
43
+ x = torch.stack([torch.from_numpy(i).permute(2, 0, 1).float().div(255) for i in imgs])[None]
44
+ rng = np.random.default_rng(0)
45
+ boosts = {}
46
+ for _ in range(60):
47
+ y, p = A.augment(x, rng)
48
+ boosts[p["boost"]] = boosts.get(p["boost"], 0) + 1
49
+ assert y.shape == x.shape and torch.isfinite(y).all() and float(y.min()) >= 0 and float(y.max()) <= 1
50
+ out["aug_boost_counts_60"] = boosts
51
+ out["mean_abs_change_last"] = round(float((y - x).abs().mean()), 4)
52
+ q = np.zeros((1, 29))
53
+ dq = A.perturb_joints(q, rng)
54
+ out["proprio_max_abs_deg"] = round(float(np.degrees(np.abs(dq).max())), 3)
55
+ out["patched"] = Gr00tN1d7Processor._get_vlm_inputs.__name__ == "_get_vlm_inputs" and \
56
+ Gr00tN1d7Processor.__call__.__module__ == "hb_launch"
57
+ print("CPU_VERIFY", json.dumps(out), flush=True)
a14-unscrew-bulb/v2-b512/step-004000/config.json ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_horizon": 40,
3
+ "add_pos_embed": true,
4
+ "apply_sincos_state_encoding": false,
5
+ "architectures": [
6
+ "Gr00tN1d7"
7
+ ],
8
+ "attn_dropout": 0.2,
9
+ "attn_implementation": null,
10
+ "backbone_embedding_dim": 2048,
11
+ "backbone_trainable_params_fp32": true,
12
+ "color_jitter_params": {
13
+ "brightness": 0.3,
14
+ "contrast": 0.4,
15
+ "hue": 0.08,
16
+ "saturation": 0.5
17
+ },
18
+ "crop_fraction": 0.95,
19
+ "diffusion_model_cfg": {
20
+ "attention_head_dim": 48,
21
+ "dropout": 0.2,
22
+ "final_dropout": true,
23
+ "interleave_self_attention": true,
24
+ "norm_type": "ada_norm",
25
+ "num_attention_heads": 32,
26
+ "num_layers": 32,
27
+ "output_dim": 1024,
28
+ "positional_embeddings": null
29
+ },
30
+ "dtype": "bfloat16",
31
+ "exclude_state": false,
32
+ "formalize_language": true,
33
+ "hidden_size": 1024,
34
+ "image_crop_size": [
35
+ 230,
36
+ 230
37
+ ],
38
+ "image_target_size": [
39
+ 256,
40
+ 256
41
+ ],
42
+ "letter_box_transform": false,
43
+ "load_bf16": false,
44
+ "max_action_dim": 132,
45
+ "max_num_embodiments": 32,
46
+ "max_seq_len": 1024,
47
+ "max_state_dim": 132,
48
+ "model_dtype": "bfloat16",
49
+ "model_name": "nvidia/Cosmos-Reason2-2B",
50
+ "model_type": "Gr00tN1d7",
51
+ "noise_beta_alpha": 1.5,
52
+ "noise_beta_beta": 1.0,
53
+ "noise_s": 0.999,
54
+ "num_inference_timesteps": 4,
55
+ "num_timestep_buckets": 1000,
56
+ "random_history_crop": true,
57
+ "random_rotation_angle": 0,
58
+ "reproject_vision": false,
59
+ "rtc_ramp_rate": 6.0,
60
+ "select_layer": 16,
61
+ "shortest_image_edge": 256,
62
+ "state_dropout_prob": 0.2,
63
+ "state_gaussian_noise_std": 0.0,
64
+ "strict_action_padding_mask": false,
65
+ "strict_vl_padding_mask": false,
66
+ "transformers_version": "4.57.3",
67
+ "tune_diffusion_model": true,
68
+ "tune_linear": true,
69
+ "tune_llm": true,
70
+ "tune_projector": true,
71
+ "tune_top_llm_layers": 0,
72
+ "tune_visual": true,
73
+ "tune_vlln": true,
74
+ "use_albumentations": true,
75
+ "use_alternate_vl_dit": true,
76
+ "use_flash_attention": true,
77
+ "use_future_tokens": false,
78
+ "use_mean_std": false,
79
+ "use_percentiles": true,
80
+ "use_vl_self_attention": true,
81
+ "use_vlln": true,
82
+ "vl_self_attention_cfg": {
83
+ "attention_head_dim": 64,
84
+ "dropout": 0.2,
85
+ "final_dropout": true,
86
+ "num_attention_heads": 32,
87
+ "num_layers": 4,
88
+ "positional_embeddings": null
89
+ }
90
+ }
a14-unscrew-bulb/v2-b512/step-004000/dataset_meta/hb-dataset-receipt.json ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "a14v2",
3
+ "source": "/mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/a14-unscrew-bulb-v2-8411a6d/data",
4
+ "episodes": 100,
5
+ "frames": 185638,
6
+ "video_keys": [
7
+ "observation.images.head",
8
+ "observation.images.chest"
9
+ ],
10
+ "state_dim_raw": 75,
11
+ "body_index": [
12
+ 0,
13
+ 1,
14
+ 2,
15
+ 3,
16
+ 4,
17
+ 5,
18
+ 6,
19
+ 7,
20
+ 8,
21
+ 9,
22
+ 10,
23
+ 11,
24
+ 12,
25
+ 13,
26
+ 14,
27
+ 15,
28
+ 16,
29
+ 17,
30
+ 18,
31
+ 19,
32
+ 20,
33
+ 21,
34
+ 22,
35
+ 23,
36
+ 24,
37
+ 25,
38
+ 26,
39
+ 27,
40
+ 28,
41
+ 29,
42
+ 30
43
+ ],
44
+ "stats_tag": "NEW_EMBODIMENT",
45
+ "modality_sha256": "4969468a30008415555bd168ad500c64011f3e0a47718724d1318849d4e162fa",
46
+ "stats_sha256": "4be3342282bad2738d3dc1305154d06c47fb31b86e75e1fbd4a29b49c2d211a8"
47
+ }
a14-unscrew-bulb/v2-b512/step-004000/dataset_meta/modality.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "state": {
3
+ "left_leg": {
4
+ "original_key": "observation.state",
5
+ "start": 0,
6
+ "end": 6
7
+ },
8
+ "right_leg": {
9
+ "original_key": "observation.state",
10
+ "start": 6,
11
+ "end": 12
12
+ },
13
+ "waist": {
14
+ "original_key": "observation.state",
15
+ "start": 12,
16
+ "end": 15
17
+ },
18
+ "left_arm": {
19
+ "original_key": "observation.state",
20
+ "start": 15,
21
+ "end": 22
22
+ },
23
+ "right_arm": {
24
+ "original_key": "observation.state",
25
+ "start": 22,
26
+ "end": 29
27
+ },
28
+ "head": {
29
+ "original_key": "observation.state",
30
+ "start": 29,
31
+ "end": 31
32
+ },
33
+ "left_hand": {
34
+ "original_key": "observation.state",
35
+ "start": 31,
36
+ "end": 53
37
+ },
38
+ "right_hand": {
39
+ "original_key": "observation.state",
40
+ "start": 53,
41
+ "end": 75
42
+ },
43
+ "projected_gravity": {
44
+ "original_key": "observation.projected_gravity",
45
+ "start": 0,
46
+ "end": 3
47
+ }
48
+ },
49
+ "action": {
50
+ "body_target": {
51
+ "original_key": "action.body_target",
52
+ "start": 0,
53
+ "end": 31
54
+ },
55
+ "left_hand_target": {
56
+ "original_key": "action.hands",
57
+ "start": 0,
58
+ "end": 22
59
+ },
60
+ "right_hand_target": {
61
+ "original_key": "action.hands",
62
+ "start": 22,
63
+ "end": 44
64
+ },
65
+ "planner": {
66
+ "original_key": "action.planner",
67
+ "start": 0,
68
+ "end": 7
69
+ }
70
+ },
71
+ "video": {
72
+ "head": {
73
+ "original_key": "observation.images.head"
74
+ },
75
+ "chest": {
76
+ "original_key": "observation.images.chest"
77
+ }
78
+ },
79
+ "annotation": {
80
+ "human.task_description": {
81
+ "original_key": "task_index"
82
+ }
83
+ }
84
+ }
a14-unscrew-bulb/v2-b512/step-004000/embodiment_id.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "robocasa_panda_omron": 13,
3
+ "oxe_droid": 17,
4
+ "oxe_fractal": 18,
5
+ "oxe_language_table": 19,
6
+ "oxe_bridge": 20,
7
+ "unknown": 22,
8
+ "gr1_unified": 20,
9
+ "agibot": 26,
10
+ "sim_behavior_r1_pro": 23,
11
+ "xdof": 24,
12
+ "xdof_oss_data": 25,
13
+ "unitree_g1_full_body_with_waist_height_nav_cmd": 25,
14
+ "real_r1_pro_sharpa": 27,
15
+ "real_r1_pro_sharpa_add_view": 27,
16
+ "real_r1_pro_sharpa_relative_arm_joint": 26,
17
+ "real_r1_pro_sharpa_delta_eef": 26,
18
+ "real_r1_pro_sharpa_absolute_eef": 26,
19
+ "real_r1_pro_sharpa_meanstd": 26,
20
+ "real_r1_pro_sharpa_relative_eef": 26,
21
+ "real_r1_pro_sharpa_relative_eef_add_view": 26,
22
+ "real_r1_pro_sharpa_relative_eef_relative_hand": 26,
23
+ "real_r1_pro_sharpa_relative_eef_human": 26,
24
+ "real_r1_pro_sharpa_relative_eef_human_add_view": 26,
25
+ "real_r1_pro_sharpa_relative_eef_human_relative_hand": 26,
26
+ "real_r1_pro_sharpa_relative_eef_egodex": 26,
27
+ "real_r1_pro_sharpa_relative_eef_egodex_relative_hand": 26,
28
+ "real_r1_pro_sharpa_relative_eef_egodex_wrist_only": 26,
29
+ "real_r1_pro_sharpa_relative_eef_maxinsights": 26,
30
+ "real_r1_pro_sharpa_relative_eef_maxinsights_relative_hand": 26,
31
+ "real_r1_pro_sharpa_relative_eef_mecka": 26,
32
+ "real_r1_pro_sharpa_relative_eef_mecka_relative_hand": 26,
33
+ "real_g1_relative_eef_absolute_joints": 25,
34
+ "real_g1_relative_eef_absolute_joints_wrist_cam": 25,
35
+ "real_g1_relative_eef_relative_joints": 25,
36
+ "real_r1_pro_sharpa_relative_eef_relative_hand_relative_joint": 26,
37
+ "real_r1_pro_sharpa_relative_joint": 29,
38
+ "oxe_droid_relative_eef_relative_joint": 24,
39
+ "oxe_droid_relative_eef_relative_joint_swapped": 24,
40
+ "oxe_droid_relative_eef_relative_joint_upweight_z": 24,
41
+ "oxe_droid_relative_eef_relative_joint_upweight_z_swapped": 24,
42
+ "oxe_droid_relative_eef_relative_joint_3view": 24,
43
+ "oxe_droid_relative_eef_relative_joint_3view_swapped": 24,
44
+ "oxe_droid_relative_eef": 24,
45
+ "oxe_droid_joint_position_relative": 24,
46
+ "xdof_relative_eef_relative_joint": 27,
47
+ "xdof_relative_eef_relative_joint_subtask": 27,
48
+ "xdof_relative_eef": 27,
49
+ "xdof_relative_joint": 28,
50
+ "simpler_env_google": 0,
51
+ "simpler_env_widowx": 1,
52
+ "libero_sim": 2,
53
+ "droid_sim": 3,
54
+ "robocasa_gr1_tabletop": 10,
55
+ "new_embodiment": 10,
56
+ "unitree_g1_sonic": 11
57
+ }
a14-unscrew-bulb/v2-b512/step-004000/experiment_cfg/conf.yaml ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ load_config_path: null
2
+ model:
3
+ model_type: Gr00tN1d7
4
+ model_dtype: bfloat16
5
+ model_name: nvidia/Cosmos-Reason2-2B
6
+ backbone_model_type: qwen
7
+ model_revision: null
8
+ tune_top_llm_layers: 0
9
+ backbone_embedding_dim: 2048
10
+ tune_llm: true
11
+ tune_visual: true
12
+ select_layer: 12
13
+ reproject_vision: false
14
+ use_flash_attention: true
15
+ load_bf16: false
16
+ backbone_trainable_params_fp32: true
17
+ image_crop_size:
18
+ - 230
19
+ - 230
20
+ image_target_size:
21
+ - 256
22
+ - 256
23
+ shortest_image_edge: null
24
+ crop_fraction: null
25
+ random_rotation_angle: null
26
+ color_jitter_params: null
27
+ use_albumentations_transforms: true
28
+ letter_box_transform: false
29
+ extra_augmentation_config: null
30
+ formalize_language: true
31
+ apply_sincos_state_encoding: false
32
+ use_percentiles: true
33
+ use_relative_action: true
34
+ max_state_dim: 132
35
+ max_action_dim: 132
36
+ action_horizon: 40
37
+ strict_action_padding_mask: null
38
+ hidden_size: 1024
39
+ input_embedding_dim: 1536
40
+ state_history_length: 1
41
+ add_pos_embed: true
42
+ attn_dropout: 0.2
43
+ use_vlln: true
44
+ strict_vl_padding_mask: null
45
+ max_seq_len: 1024
46
+ use_alternate_vl_dit: true
47
+ attend_text_every_n_blocks: 2
48
+ diffusion_model_cfg:
49
+ positional_embeddings: null
50
+ num_layers: 16
51
+ num_attention_heads: 32
52
+ attention_head_dim: 48
53
+ norm_type: ada_norm
54
+ dropout: 0.2
55
+ final_dropout: true
56
+ output_dim: 1024
57
+ interleave_self_attention: true
58
+ num_inference_timesteps: 4
59
+ noise_beta_alpha: 1.5
60
+ noise_beta_beta: 1.0
61
+ noise_s: 0.999
62
+ num_timestep_buckets: 1000
63
+ tune_projector: true
64
+ tune_diffusion_model: true
65
+ tune_vlln: true
66
+ state_dropout_prob: 0.2
67
+ exclude_state: false
68
+ use_mean_std: false
69
+ max_num_embodiments: 32
70
+ data:
71
+ datasets:
72
+ - dataset_paths:
73
+ - /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-data/a14v2
74
+ embodiment_tag: new_embodiment
75
+ mix_ratio: 1.0
76
+ dataset_type: physical_embodiment
77
+ val_dataset_path: null
78
+ modality_configs:
79
+ new_embodiment:
80
+ video:
81
+ delta_indices:
82
+ - 0
83
+ modality_keys:
84
+ - head
85
+ - chest
86
+ sin_cos_embedding_keys: null
87
+ mean_std_embedding_keys: null
88
+ action_configs: null
89
+ state:
90
+ delta_indices:
91
+ - 0
92
+ modality_keys:
93
+ - left_leg
94
+ - right_leg
95
+ - waist
96
+ - left_arm
97
+ - right_arm
98
+ - head
99
+ - left_hand
100
+ - right_hand
101
+ - projected_gravity
102
+ sin_cos_embedding_keys: null
103
+ mean_std_embedding_keys: null
104
+ action_configs: null
105
+ action:
106
+ delta_indices:
107
+ - 0
108
+ - 1
109
+ - 2
110
+ - 3
111
+ - 4
112
+ - 5
113
+ - 6
114
+ - 7
115
+ - 8
116
+ - 9
117
+ - 10
118
+ - 11
119
+ - 12
120
+ - 13
121
+ - 14
122
+ - 15
123
+ modality_keys:
124
+ - body_target
125
+ - left_hand_target
126
+ - right_hand_target
127
+ - planner
128
+ sin_cos_embedding_keys: null
129
+ mean_std_embedding_keys: null
130
+ action_configs:
131
+ - rep: ABSOLUTE
132
+ type: NON_EEF
133
+ format: DEFAULT
134
+ state_key: null
135
+ - rep: ABSOLUTE
136
+ type: NON_EEF
137
+ format: DEFAULT
138
+ state_key: null
139
+ - rep: ABSOLUTE
140
+ type: NON_EEF
141
+ format: DEFAULT
142
+ state_key: null
143
+ - rep: ABSOLUTE
144
+ type: NON_EEF
145
+ format: DEFAULT
146
+ state_key: null
147
+ language:
148
+ delta_indices:
149
+ - 0
150
+ modality_keys:
151
+ - annotation.human.task_description
152
+ sin_cos_embedding_keys: null
153
+ mean_std_embedding_keys: null
154
+ action_configs: null
155
+ download_cache: false
156
+ shard_size: 1024
157
+ episode_sampling_rate: 0.1
158
+ num_shards_per_epoch: 100000
159
+ ds_weights_alpha: null
160
+ override_pretraining_statistics: true
161
+ mode: single_turn
162
+ random_chop: 0.0
163
+ mock_dataset_mode: false
164
+ shuffle: true
165
+ seed: 42
166
+ multiprocessing_context: fork
167
+ allow_padding: false
168
+ subsample_ratio: 1.0
169
+ loader: fast
170
+ fast_loader:
171
+ decode_threads: 2
172
+ ffmpeg_threads: 1
173
+ shuffle_buffer: 0
174
+ min_fill: 0
175
+ image_crop_size:
176
+ - 244
177
+ - 244
178
+ image_target_size:
179
+ - 224
180
+ - 224
181
+ training:
182
+ output_dir: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/groot-runs-2x32/a14v2/production
183
+ experiment_name: null
184
+ max_steps: 4000
185
+ global_batch_size: 512
186
+ per_gpu_batch_size: null
187
+ gradient_accumulation_steps: 1
188
+ learning_rate: 0.0001
189
+ lr_scheduler_type: cosine
190
+ weight_decay: 1.0e-05
191
+ warmup_ratio: 0.05
192
+ warmup_steps: 0
193
+ max_grad_norm: 1.0
194
+ optim: adamw_torch
195
+ start_from_checkpoint: /mnt/nas/cross/Team/RobotAgent/Personal/baijiaqi/humanoid-benchmark-work/preflight-20261104T1648Z/policies-20261109/models/groot-n17-base
196
+ skip_weight_loading: false
197
+ tf32: true
198
+ fp16: false
199
+ bf16: true
200
+ eval_bf16: true
201
+ logging_steps: 10
202
+ save_steps: 500
203
+ save_total_limit: 5
204
+ save_vl_model: false
205
+ save_only_model: false
206
+ resume_from_checkpoint: true
207
+ upload_checkpoints: false
208
+ upload_every: 1000
209
+ upload_last_n_checkpoints: 5
210
+ max_concurrent_uploads: 2
211
+ eval_strategy: 'no'
212
+ eval_steps: 500
213
+ eval_set_split_ratio: 0.1
214
+ eval_batch_size: 2
215
+ save_best_eval_metric_name: ''
216
+ save_best_eval_metric_greater_is_better: true
217
+ deepspeed_stage: 2
218
+ gradient_checkpointing: false
219
+ transformers_trust_remote_code: true
220
+ transformers_local_files_only: false
221
+ transformers_cache_dir: null
222
+ transformers_access_token: null
223
+ use_ddp: false
224
+ ddp_bucket_cap_mb: 100
225
+ num_gpus: 16
226
+ dataloader_num_workers: 8
227
+ remove_unused_columns: false
228
+ use_wandb: false
229
+ wandb_project: finetune-gr00t-n1d7
230
+ enable_profiling: false
231
+ max_retries: 3
232
+ assert_loss_less_than: null
233
+ add_rl_callback: false
234
+ enable_open_loop_eval: false
235
+ open_loop_eval_traj_ids:
236
+ - 0
237
+ open_loop_eval_steps_per_traj: 100
238
+ open_loop_eval_plot_indices: null
239
+ max_steps: 4000
240
+ save_steps: 500