auto-sync 2026-07-02T14:08:40Z workspace

workspace/docs/hf_sync.md

@@ -0,0 +1,40 @@
+# Hugging Face Sync
+
+Target repo: `anhtld/vla`.
+
+One-shot push:
+
+```bash
+PATH="$PWD/.venv/bin:$PATH" HF_REPO_ID=anhtld/vla scripts/hf_push_once.sh
+```
+
+Continuous push every 15 minutes:
+
+```bash
+setsid env PATH="$PWD/.venv/bin:$PATH" \
+  HF_REPO_ID=anhtld/vla \
+  HF_REPO_TYPE=model \
+  HF_SYNC_INTERVAL_SECONDS=900 \
+  scripts/hf_push_every_15m.sh \
+  >> outputs/hf_sync/hf_push_daemon_login.out 2>&1 < /dev/null &
+echo $! > outputs/hf_sync/hf_push_daemon_login.pid
+```
+
+Stop the login-node daemon:
+
+```bash
+kill "$(cat outputs/hf_sync/hf_push_daemon_login.pid)"
+```
+
+Monitor:
+
+```bash
+tail -f outputs/hf_sync/hf_push_daemon_login.out
+tail -f outputs/hf_sync/hf_sync.log
+```
+
+The sync excludes `.env`, token-like files, keys, `.git`, virtualenvs, containers,
+and build caches. It includes code, `latex/`, results, logs, outputs, checkpoints,
+selected scratch experiment roots, and cache roots configured in
+`scripts/hf_push_once.sh`.
+

workspace/dovla_cil/eval/maniskill_policy_rollout.py

@@ -111,6 +111,7 @@ def evaluate_maniskill_policy_rollout(
     lattice_exclude_types: tuple[str, ...] = (),
     lattice_exclude_type_tasks: dict[str, tuple[str, ...]] | None = None,
     candidate_type_bonuses: dict[str, float] | None = None,
+    candidate_type_bonuses_by_task: dict[str, dict[str, float]] | None = None,
     candidate_type_bonus_components: bool = False,
     field_rank_biases_by_task: dict[str, list[float]] | None = None,
     candidate_oracle_rollouts: int = 0,
@@ -294,6 +295,13 @@ def evaluate_maniskill_policy_rollout(
         str(candidate_type): float(bonus)
         for candidate_type, bonus in (candidate_type_bonuses or {}).items()
     }
+    candidate_type_bonuses_by_task = {
+        str(task_id): {
+            str(candidate_type): float(bonus)
+            for candidate_type, bonus in bonuses.items()
+        }
+        for task_id, bonuses in (candidate_type_bonuses_by_task or {}).items()
+    }
     field_rank_biases_by_task = {
         str(task_id): [float(value) for value in values]
         for task_id, values in (field_rank_biases_by_task or {}).items()
@@ -412,6 +420,11 @@ def evaluate_maniskill_policy_rollout(
     task_summaries: dict[str, dict[str, Any]] = {}
     with torch.no_grad():
         for task_id, task_cases in sorted(by_task.items()):
+            task_candidate_type_bonuses = _task_candidate_type_bonuses(
+                candidate_type_bonuses,
+                candidate_type_bonuses_by_task,
+                task_id=task_id,
+            )
             task_rows = _evaluate_task_cases(
                 task_id,
                 task_cases,
@@ -458,7 +471,7 @@ def evaluate_maniskill_policy_rollout(
                 ),
                 lattice_exclude_types=lattice_exclude_types,
                 lattice_exclude_type_tasks=lattice_exclude_type_tasks,
-                candidate_type_bonuses=candidate_type_bonuses,
+                candidate_type_bonuses=task_candidate_type_bonuses,
                 candidate_type_bonus_components=candidate_type_bonus_components,
                 field_rank_biases=(
                     field_rank_biases_by_task.get(task_id)
@@ -624,6 +637,7 @@ def evaluate_maniskill_policy_rollout(
             for candidate_type, tasks in lattice_exclude_type_tasks.items()
         },
         "candidate_type_bonuses": candidate_type_bonuses,
+        "candidate_type_bonuses_by_task": candidate_type_bonuses_by_task,
         "candidate_type_bonus_components": bool(candidate_type_bonus_components),
         "field_rank_biases_by_task": field_rank_biases_by_task,
         "candidate_oracle_rollouts": int(candidate_oracle_rollouts),
@@ -3112,6 +3126,18 @@ def _candidate_type_matches_exclusion(candidate_type: str, excluded: set[str]) -
     return any(part in excluded for part in candidate_type.split("+"))
 
 
+def _task_candidate_type_bonuses(
+    global_bonuses: dict[str, float],
+    bonuses_by_task: dict[str, dict[str, float]],
+    *,
+    task_id: str,
+) -> dict[str, float]:
+    bonuses = dict(global_bonuses)
+    bonuses.update(bonuses_by_task.get("*", {}))
+    bonuses.update(bonuses_by_task.get(task_id, {}))
+    return bonuses
+
+
 def _lattice_candidate_type_bonus(
     batch: list[_RolloutCase],
     *,

workspace/latex/main.log

@@ -1,4 +1,4 @@
-This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  2 JUL 2026 09:28
+This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  2 JUL 2026 10:04
 entering extended mode
  restricted \write18 enabled.
  %&-line parsing enabled.
@@ -494,11 +494,11 @@ Package rerunfilecheck Info: File `main.out' has not changed.
 Here is how much of TeX's memory you used:
  9876 strings out of 480884
  146359 string characters out of 5900692
- 453089 words of memory out of 5000000
+ 454089 words of memory out of 5000000
  26968 multiletter control sequences out of 15000+600000
  412837 words of font info for 63 fonts, out of 8000000 for 9000
  36 hyphenation exceptions out of 8191
- 71i,8n,74p,308b,335s stack positions out of 5000i,500n,10000p,200000b,80000s
+ 71i,8n,74p,308b,393s stack positions out of 5000i,500n,10000p,200000b,80000s
 {/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/font
 s/enc/dvips/cm-super/cm-super-ts1.enc}</cvmfs/soft.computecanada.ca/gentoo/2023
 /x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb></cvm
@@ -521,7 +521,7 @@ e1/public/amsfonts/cm/cmr9.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64
 .computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public
 /amsfonts/cm/cmsy7.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/
 share/texmf-dist/fonts/type1/public/cm-super/sfrm1000.pfb>
-Output written on main.pdf (3 pages, 165505 bytes).
+Output written on main.pdf (3 pages, 165984 bytes).
 PDF statistics:
  118 PDF objects out of 1000 (max. 8388607)
  98 compressed objects within 1 object stream

workspace/latex/main.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b046fd725f3d9c662b58390ccdb4ee75ecfbde452737176bb1efa6dbdcd9e1cf
-size 165505
+oid sha256:3e7ce55c19c43e3c61971e97798bf2d886218188b5d2ea5bb1311718b14d525d
+size 165984

workspace/latex/main.tex

@@ -34,6 +34,9 @@ manipulation tasks, direct h=16 behavior cloning reaches 29.74\% success,
 whereas the best deployment-clean transported residual field reaches 38.84\%
 without same-state validation proposals or expert proposals. Same-state
 no-expert lattices reach 56.99\%, exposing a remaining proposal-generation gap.
+Within the current clean proposal prefix, a diagnostic top-8 candidate oracle
+reaches 44.35\%, showing that conditional selector calibration still leaves
+measurable headroom.
 Negative ablations show that simply increasing measured support, using
 rank-only calibration, or optimizing the field off-manifold does not explain
 the gain; the strongest clean result comes from a tiny train-source utility
@@ -135,7 +138,11 @@ transported residual field reaches 38.84\%. This gap is the paper's most useful
 tension: \cil{} supervision reveals a strong local selector, but deployment
 still needs better proposal generation. The clean gain is nonetheless large:
 the best clean row improves over direct h=16 behavior cloning by 9.10
-percentage points.
+percentage points. A diagnostic candidate oracle over the top-8 clean proposal
+prefix for this best row reaches 44.35\%, with mean best branch rank 2.48 and a
+mean oracle score gain of 0.107 over the selected branch. This keeps the
+proposal source deployment-clean and isolates selector headroom rather than
+relying on same-state validation candidates.
 
 Several negative results sharpen the claim. B24 measured support reaches only
 38.61\%, below the B12 exact-support row. Rank-only calibration reaches 38.72\%,
@@ -157,4 +164,3 @@ yet justify an unqualified SOTA claim.
 \bibliographystyle{plain}
 \bibliography{references}
 \end{document}
-

workspace/latex/tables/main_results.tex

@@ -15,10 +15,10 @@ K4 compatible tangent near-miss challenger & yes & 36.06 & +6.32 \\
 K6 transported residual field, exact support & yes & 38.78 & +9.04 \\
 K6 transported residual field + source-score 0.01 & yes & \textbf{38.84} & \textbf{+9.10} \\
 \midrule
+K6 source-score top-8 candidate oracle & diagnostic & 44.35 & +14.61 \\
 Same-state no-expert lattice & diagnostic & 56.99 & +27.25 \\
 Same-state full lattice & diagnostic & 69.33 & +39.59 \\
 Oracle ceiling & diagnostic & 86.78 & +57.04 \\
 \bottomrule
 \end{tabular}
 \end{table}
-