t2ance/mat-05-persistent-buffer-and-hint-ceiling

05 Persistent buffer and hint ceiling

With a persistent cross-round tree buffer (an AceGRPO-style evolving buffer, commit 1ca9b09) added on top of tree-GRPO's learnability-based parent selection, does the valid_auc_max plateau on spooky-author-identification (stuck at -0.345 to -0.365 in an earlier fresh-tree run) move, and if it moves once the task's modeling hint is relaxed, does that movement belong to the hint's prescribed recipe or to the Qwen3.5-27B policy's own capability? This repository is the data root of that question: everything its five runs wrote, minus the exclusion list in section 5.

1. Question and links

• Report: personal/experiments/rl-training/05-persistent-buffer-and-hint-ceiling/report/main.tex in t2ance/ml-agent-training-personal, opened 2026-07-12, related to reports 01, 04, 06 and 08 of the same rl-training line (per the report's own header); no card.md exists for any of the three leaves (tree_grpo_27b_spooky_buffer, tree_grpo_27b_spooky_buffer_minhint, tree_grpo_27b_spooky_stable) — each leaf's configs/run_*.sh (in the question directory, not this repository) holds the launcher and its pre-launch header, and the report's own Evidence paragraphs are the run-to-leaf mapping used below.
• Issue: none named by the report or by any leaf's launcher.
• Answer, as the report's Conclusion states it: the buffer mechanism works ("from round 2 the learnability selection draws on real child evidence", frac_selections_with_evidence nonzero in 16 of 20 rounds, versus 16 of 18 selections on the uninformative prior in report 04's fresh-tree run), but under the strong hint it did not move the ceiling (valid_auc_max stayed within -0.3641 to -0.3448 for 21 steps); once the hint was reduced to hard constraints and the timeout raised to 900s, the score crossed the strong-hint reference ceiling -0.3448 once, at -0.3156 (step 28, +0.0292 by this root's own full-precision arithmetic on the round's best node against the reference ceiling -- the report states +0.0293 for the same crossing in five places, main.tex lines 42, 48, 465, 481, 562, a figure this README does not reproduce from the round's own .jsonl record, section 3), through a TF-IDF plus LogisticRegression plus ComplementNB blend the strong hint had excluded, and fell back to -0.3474 the next round; the report's verdict is: "The question of this report is therefore not answered; what is answered is that the buffer mechanism operates and that the strong hint's band was not crossed while the hint was in force."
• W&B project: pqin/tree-grpo. Runs: tree_grpo_27b_spooky_buffer (Exp. 1, leaf tree_grpo_27b_spooky_buffer, 21 history rows), tree_grpo_27b_spooky_buffer_minhint (Exp. 2, leaf tree_grpo_27b_spooky_buffer_minhint, 4 rows), tree_grpo_27b_spooky_buffer_minhint_t900 (Exp. 3, same leaf, 4 rows), tree_grpo_27b_spooky_stable (Exp. 4 first attempt, leaf tree_grpo_27b_spooky_stable, crashed before step 1, no history rows), tree_grpo_27b_spooky_stable2 (Exp. 4 second attempt, same leaf, 2 rows).
• Status: finished — all five runs have ended, none are live; the campaign ran from 2026-07-12T07:14Z (Exp. 1's launch) to 2026-07-13 (Exp. 4's second attempt, killed by Ray on host RAM while round 3 was already graded; its files are dated Jul 13 22:24 local), and the report treats the arc as closed with the underlying question still open, naming a frozen-actor control on spooky as future work.

2. Directory tree

05-persistent-buffer-and-hint-ceiling/
├── README.md                                                        # this front page
├── 20260712_064500/                                                 # Exp. 1 (report Section 4.1 table): tree_grpo_27b_spooky_buffer, strong hint, fresh weights + empty buffer, 300s timeout, launched 2026-07-12T07:14Z, killed at host RAM 223 GB after round 21's training step
│   ├── round_1/
│   │   ├── spooky-author-identification__1783840449.jsonl            # 32 node records generated this round (row fields: section 4)
│   │   ├── spooky-author-identification__1783840449__transcript.md   # the same round as a human-readable Markdown rendering
│   │   ├── spooky-author-identification__1783840449__tree.txt        # the same round as a one-line-per-node plain-text tree
│   │   └── spooky-author-identification__summary_1783841724.json     # the round's summary row (row fields: section 4)
│   ├── round_2/ .. round_20/                                         # 19 more rounds, same 4-file shape as round_1 (timestamps differ); step 20 is the run's last saved checkpoint (checkpoint itself not present, section 5)
│   ├── stale_round_21/                                              # round 21: rollout finalized (32 nodes, 23 valid) and trained -- its summary_*.json shows phase "train"; its training row is timing.jsonl's last (step 21), matching W&B step 21, but no checkpoint was saved for this step (step 20's remains the run's last saved checkpoint)
│   ├── stale_round_22/                                              # round 22: rollout never finalized (only 8 of 32 node records written, no summary_*.json file) and never trained -- the host-RAM kill landed mid-rollout, just after round 21's training step had already completed
│   ├── timing.jsonl                                                  # 21 rows, steps 1-21: per-step harness phase timing (row fields: section 4)
│   ├── viz_dashboard_spooky-author-identification.png                # the run's rendered dashboard (harness/viz.py, commit d02ef3e)
│   └── viz_tree_spooky-author-identification.png                     # the run's rendered persistent-tree map
├── 20260712_222742/                                                  # Exp. 2 (report Section 4.2 table): tree_grpo_27b_spooky_buffer_minhint, hint reduced to hard constraints, 300s timeout, resumed weights+optimizer from Exp. 1's global_step_20 with its buffer reloaded, launched 2026-07-12T23:06Z
│   ├── round_21/ .. round_25/                                        # same 4-file shape as round_1 above; round_25's rollout is finalized (32 nodes graded) but its training step never reached timing.jsonl or W&B (the process ended right after checkpointing global_step_25)
│   ├── timing.jsonl                                                  # 4 rows, steps 21-24
│   ├── viz_dashboard_spooky-author-identification.png
│   └── viz_tree_spooky-author-identification.png
├── 20260713_030633/                                                  # Exp. 3 (report Section 4.3 table): tree_grpo_27b_spooky_buffer_minhint_t900, hint still reduced, timeout raised to 900s, resumed from Exp. 2's global_step_25 with the buffer reloaded from both prior run directories, launched 2026-07-13T03:41Z
│   ├── round_26/ .. round_30/                                        # same 4-file shape; round_28 holds the campaign's best node (id b04c251ea26f, oriented -0.3155527; this root's own arithmetic gives +0.0292 over the reference ceiling, the report states +0.0293, section 1)
│   ├── timing.jsonl                                                  # 5 rows, steps 26-30 (step 30's row is here even though the W&B history for this run stops at step 29, section 3)
│   ├── viz_dashboard_spooky-author-identification.png
│   └── viz_tree_spooky-author-identification.png
├── 20260713_170125/                                                  # Exp. 4 first attempt (report Section 4.4): tree_grpo_27b_spooky_stable, fresh weights, crashed at the first backward (FSDP2 mixed-precision dtype error under bypass_mode, fixed by commit aa2fbdb) before any step trained
│   ├── round_1/                                                      # same 4-file shape; orphaned — rollout ran, no training step, no checkpoint
│   ├── viz_dashboard_spooky-author-identification.png
│   └── viz_tree_spooky-author-identification.png                    # no timing.jsonl in this run: no step ever trained
└── 20260713_180808/                                                  # Exp. 4 second attempt (report Section 4.4): tree_grpo_27b_spooky_stable2, fresh weights + empty buffer, bypass_mode stabilizers, launched 2026-07-13, killed by Ray on host RAM (247.8 of 251.6 GB) while round 3's nodes were already graded
    ├── round_1/ .. round_3/                                          # same 4-file shape; round_3 graded but never trained (the kill landed before its training step)
    ├── timing.jsonl                                                  # 2 rows, steps 1-2
    ├── viz_dashboard_spooky-author-identification.png
    └── viz_tree_spooky-author-identification.png

No trained-weight checkpoint survives on the host for any of the five runs above (report, Exp. 3 Observations: "the checkpoint directory the launcher names is no longer present on the host"; confirmed empty here for tree_grpo_qwen3_5_27b_spooky_buffer and tree_grpo_qwen3_5_27b_spooky_stable under /data2/peijia/ml-agent-checkpoints/tree-grpo/), so this data root holds only the tree-search side of each run — the round records, the harness's per-step phase timing, and the rendered diagnostics — and none of the checkpoint-shaped exclusions in section 5 finds anything to exclude. There is no logs/, manifests/, wandb/ client directory or judge_calls.jsonl under this root: the console log and the W&B API export that the report cites live in the report's own evidence directory (personal/experiments/rl-training/05-persistent-buffer-and-hint-ceiling/results/) in the ml-agent-training-personal repository, not in this data root, and this task's reward is the competition's own log-loss grading of an executed script, not an LLM judge, so no judge-call log exists to publish. No other question's report cites a run in this root as evidence: report 08 (08-bond-blind-judge) continues from the code state this arc reached (the reduced hint and the persistent buffer, both still in force at the product repository's head per this report's Future work) but launches its own fresh run rather than reusing one of the five runs above, and report 06's frozen-actor control was never run for spooky.

3. How to read each kind of file

• Small text, JSON, JSONL and image files: https://huggingface.co/datasets/t2ance/mat-05-persistent-buffer-and-hint-ceiling/raw/main/<path>, for example raw/main/20260712_064500/timing.jsonl, raw/main/20260712_064500/round_1/spooky-author-identification__summary_1783841724.json, raw/main/20260712_064500/round_1/spooky-author-identification__1783840449__tree.txt or raw/main/20260712_064500/viz_dashboard_spooky-author-identification.png.
• The per-round node-record .jsonl (it carries every node's full prompt messages, thinking, code and every sampled token's id and logprob, so it runs from 5 MB in round 1 to 7 MB by round 28) and its matching __transcript.md (a Markdown rendering of the same round that can pass 10 MB by later rounds): resolve/main/<path>, for example resolve/main/20260713_030633/round_28/spooky-author-identification__1783922088.jsonl (7.0 MB) and resolve/main/20260713_030633/round_28/spooky-author-identification__1783922088__transcript.md (10.6 MB). raw/ returns only the Git LFS pointer for a file the repository tracks with LFS; resolve/ always returns the real bytes, so it is the safe default whenever a file's LFS status is not known.
• The report's headline metrics and how to recompute them from this repository alone, independent of W&B (as the report's Verification section states every decisive number was re-read): valid_auc_max for a round is -1 * the minimum test_metric over that round's own .jsonl records with status == "ok" -- read the round's own file, never the summary field. A round's summary_*.json best_test_metric is a different quantity: the cumulative best-ever-in-this-run value, carried forward unchanged from the previous round unless the current round happens to set a new campaign-wide low, so -1 * best_test_metric equals the round's own valid_auc_max only on the rounds that set that new best, never in general. Checked across every round of 20260712_064500: rounds 1-4, 11, 13 and 19 each set a new campaign-wide best, so their own minimum test_metric does match best_test_metric (round 1: best_test_metric = 0.36405881938467716, so valid_auc_max = -0.3641, matching the report's step-1 row); every other round diverges -- for example round 5's own minimum test_metric is 0.35020554412938065 (matching W&B tree_grpo_27b_spooky_buffer step 5's rollout/valid_auc_max = -0.35020554412938065 exactly) while its summary_*.json best_test_metric still reads 0.347949604074495, the value frozen since round 4; the same freeze recurs after round 19 (rounds 20-21 both read 0.3447663084722123 in best_test_metric while their own minima are 0.3528688527428701 and 0.3497691598060371). valid_auc_mean is the mean of the oriented field over a round's own status == "ok" records; the valid rate is summary_*.json's valid / nodes. Verified directly: round 1 of 20260712_064500 has mean oriented over its 29 ok records of -0.4053948339657275 (matching the report's valid_auc_mean = -0.4054).
• parent_P and parent_P_evidence on a node instrument the report's Preliminary formula P(n) = min(sigma, 0.25) * clip(1 - mu, 0, 1) over a parent's graded children, with P(n) = 0.1 (the P_INIT prior) whenever parent_P_evidence is false; the report's frac_selections_with_evidence and parent_P_selected_mean per round are aggregates of these two fields over the round's improve/debug nodes.
• W&B under-reports two of the five runs, so this repository's own files are the complete record for them: Exp. 2's run (tree_grpo_27b_spooky_buffer_minhint) has only 4 history rows (steps 21-24) because the process ended right after checkpointing global_step_25, before logging that step, and this repository's 20260712_222742/timing.jsonl likewise stops at step 24 (the harness never completed that step's phase timing either); Exp. 3's run (..._minhint_t900) has 4 rows on W&B (steps 26-29) while 20260713_030633/timing.jsonl carries a fifth row for step 30 (its phases_s.save_checkpoint entry confirms the checkpoint save), matching the report's console-only step-30 reading. timing.jsonl's phases_s keys are the harness's own wall-clock phase timer, unrelated to and unaffected by the W&B gaps above.

4. Row fields

round_N/<task>__<timestamp>.jsonl, one row per generated tree node (32 per round: 4 expansions of 8 siblings): id, step, round, exp_id; parent_id, sibling_group (the GRPO group of 8 this node belongs to), mode (draft from the task description, improve an edit of a valid parent, debug a fix of a crashed parent), plan, depth, n_children; code, raw_completion, thinking, prompt_messages; response_token_ids, response_logprobs (the sampling engine's per-token logprobs); where (the local execution slot, local:slot0/1/2), exec_time (wall-clock seconds), exc_type (the exception class on a crash or timeout, null on a clean run), term_tail (the tail of the executed script's stdout/stderr), returncode; exec_timings, resources, gpu_util, gpu_util_peak, gpu_mem_used_gi, gpu_name, cpu_util_mean, cpu_util_peak, mem_peak_gi, kube_node, attempt (all null in every record seen: exec_backend=local, CPU-only, so these k8s/GPU-backend fields are unused here); decl_malformed, decl_deps_ignored, decl_miss_reason (declaration-parsing diagnostics, unused for this CPU-script task); is_buggy, status (ok, timeout, crashed, or no_code when the completion had no extractable code block), grade_error, test_metric (the competition's raw log loss, null unless status == "ok"), oriented (the higher-is-better view the report calls valid_auc_*, = -test_metric for a valid node), reward (the GRPO reward: oriented for a valid node, a fixed sentinel otherwise), source (grade or none), parent_P, parent_P_evidence, parent_oriented, improve_delta, group_best (whether this node is the best oriented in its sibling group), medal (the competition leaderboard tier the score reached, e.g. median, or null when it does not reach one; reference values are in the report's Background section).

round_N/<task>__summary_<timestamp>.json, one row per round: exp_id, round, model (the served weights path), exec_backend, adapter_path; nodes (32 every round), valid, expansions; buffer_nodes (the persistent tree's running total; = 32 * round in every summary read, confirming no reset across rounds within a run's own buffer); phase; best_test_metric (the cumulative best-ever-in-this-run value, not this round's own minimum test_metric; section 3), best_node_id, honest_test_metric (null in every summary read); rollout_file, workspace (the harness's own on-disk paths at run time, not the path in this repository); gen_llm_seconds, gen_exec_seconds (the round's total generation and execution wall-clock time).

round_N/<task>__<timestamp>__tree.txt, a plain-text console rendering with one line per node: [id] mode status test=<value or --> sib_grp=<sibling_group> | where exec_time, with a child indented under its parent by a leading └─.

round_N/<task>__<timestamp>__transcript.md, a Markdown rendering of the same round: a header giving the round's node/valid/best-AUC counts and the list of valid AUCs, a fenced-text tree block (the same shape as tree.txt with abbreviated ids and oriented/crashed/timeout labels), then one collapsible section per node reproducing its full prompt, response and execution output.

timing.jsonl, one row per completed training step (only steps that finished have a row): step, total_s, phases_s — a dict of harness-phase wall-clock seconds: update_weights, gen, old_log_prob, ref, adv, update_actor, and save_checkpoint on a step that also saved a checkpoint.

5. Exclusions and inputs

None of the rule's checkpoint-shaped exclusions finds anything to exclude in this root, because no checkpoint of any kind survives for any of the five runs: optimizer state, RNG state and other extra training-step state, full-model shards of a LoRA step, the critic or value model (GRPO has none), a resume pointer such as a latest-step file, and an intermediate full-model dump next to a saved adapter would all have lived under /data2/peijia/ml-agent-checkpoints/tree-grpo/tree_grpo_qwen3_5_27b_spooky_buffer/ or .../tree_grpo_qwen3_5_27b_spooky_stable/, and neither directory exists on the host (checked directly; the report's Exp. 3 Observations independently records the same absence for its own checkpoints). Everything else this campaign wrote — every round_N/ file and every run's timing.jsonl, viz_dashboard_*.png and viz_tree_*.png — goes up, and none of it is on the exclusion list.

Inputs from other experiments or from outside this project (not copied):

• The base model: every run's config points MODEL_PATH / actor_rollout_ref.model.path at /data2/peijia/models/qwen3_5_27b, the public Qwen3.5-27B weights; the report's own Background section names the policy "Qwen3.6-27B", a discrepancy this README flags rather than resolves, since the evidence in this root (the launch configs) supports only the path.
• The task: spooky-author-identification, an MLE-bench-style Kaggle competition (three-class author probabilities from short text, scored by log loss; leaderboard median 0.419, bronze 0.294 per the report's Background section) cached under this workspace's datasets/mlebench (symlinked to /data3/ruiyi/mlebench-playground/data on this host) plus its private held-out grading labels; both stay out of this repository as benchmark data.
• The verl dataloader placeholder: every run's DATA_DIR points at a gsm8k train/test parquet pair that only satisfies verl's dataloader shape requirement — the actual training signal comes from the harness's tree search on the task above, not from this parquet — and it stays out of this repository.
• No teacher or judge model: the reward is the competition's own log-loss grading of each node's executed script, not an LLM judge, so there is no external model input to name for the reward path (contrast the distillation line's ROPD runs, which do take a judge model as an input).
• The immediately preceding run tree_grpo_27b_spooky_snrfix_hint belongs to report 04's data root (04-noise-floor-and-snr-fixes, a different question) and is cited by this report only as a comparison baseline (Section 3.2, "Baseline"); none of this root's five runs resumed its weights or its buffer — Exp. 1 launched with resume_mode=disable and an empty buffer — so it is not an input under this rule, matching the empty inputs recorded for this question.