t2ance/mat-09-multitask-humanrank

09 Multitask humanrank reward and the response-length cap

Trained together under one shared humanrank reward (the percentile of a valid submission against a competition's real Kaggle leaderboard), do four heterogeneous MLE-bench tasks give the 27B LoRA policy a multi-task learning signal that no single-task run in this line had shown, and, once the first arm's response-length crash loop appeared, does raising the response-token cap rescue the missing solutions? The report's own answer (abstract and Conclusion): the shared scale kept 3 or 4 of the 4 task groups reward-spread on 23 of the 26 logged steps, but no task's valid_reward_mean or draft_reward_mean rose net-positive inside any arm, the one application of the pre-registered decision rule (step 9 of hr_b) read 0 of 4 tasks moved, and the pre-registered step-16 and step-30 gates were never reached, so part (a) was neither confirmed nor falsified by its own rule; part (b) was settled in the negative for the cap alone — raising the response cap from 12288 to 16384 rescued no text-normalization solution, since every clipped node sat exactly at the new cap — while GPU memory, not the reward signal, ended up bounding the cap (45056 total tokens OOMed in the update outright, 36864 OOMed only on the lineage's 14th update after 13 clean ones); a shared random.Random(0) seed had also made fresh drafts structurally impossible after round 1 in every arm but the last, so most of the quality reads above are composition-confounded, and the series ended on 2026-07-21 by a scope pivot to teacher SFT (distillation 02), not by the falsification rule. This repository is the data root of that question: everything its runs wrote, minus the exclusion list in section 5.

1. Question and links

• Question, cards and configs: personal/experiments/rl-training/09-multitask-humanrank/ in t2ance/ml-agent-training-personal — one card.md and one configs/run_tree_grpo_27b_mix4_<arm>.sh per launched leaf (tree_grpo_27b_mix4_hr, _hr16k, _hr24k, _hr24k_u6, _hr24k_u6b, _hr24k_u6c, _hr32k); the eighth leaf, tree_grpo_27b_mix4_hr_frozen, has only a configs/ script (the pre-registered frozen-actor control) and was never launched, so it has no card and no directory in this repository. results/ in the same tree carries the pre-audit (preaudit_multitask_20260719_final.txt), the eight runs' W&B export copies (wandb_exports/<run>/config.json, history.jsonl, summary.json), one OOM log copy, and a prompt-length figure — none of it is inside this data root; it is cited by this GitHub path, not copied here (section 5).
• Report source: personal/experiments/rl-training/09-multitask-humanrank/report/main.tex in t2ance/ml-agent-training-personal ("09. Multi-task training on a shared leaderboard-rank reward, and the response-length cap"; opened 2026-07-19, status Open in its own header — the report itself stays open for possible future work, section 5, even though the runs below are finished).
• Issue: none is named by any card or by the report.
• Code: ruz048/ml-agent-training, commit 899d203 ("harness: add ranked rewards and rollout evidence"), which the report cites as its current tracked source for harness/reward.py (the humanrank reward), harness/prompt.py and harness/search.py (the extraction contract), harness/tree.py (the per-tree seed fix), training_verl/tree_rollout.py (the per-task instrument panels) and training_verl/tree_grpo_trainer.py (the grpo_valid_only mean-only advantage); no PR is named — the harness changes ran from the working tree during the window and were committed directly.
• W&B project: pqin/tree-grpo. All eight runs recorded state: crashed in personal/docs/wandb_index.csv (every arm ended by a kill, an OOM, or a pre-rollout crash, never a clean stop): tree_grpo_27b_mix4_hr (2026-07-19T04:41Z, 0 history rows — burned by a pre-rollout batch-divisibility assertion), tree_grpo_27b_mix4_hr_b (2026-07-19T05:02Z, 10 rows), tree_grpo_27b_mix4_hr16k (2026-07-19T16:52Z, 3 rows), tree_grpo_27b_mix4_hr32k (2026-07-19T22:56Z, 0 rows — died before any step), tree_grpo_27b_mix4_hr24k (2026-07-20T02:47Z, 0 rows), tree_grpo_27b_mix4_hr24k_u6 (2026-07-20T04:54Z, 8 rows), tree_grpo_27b_mix4_hr24k_u6b (2026-07-20T16:46Z, 5 rows), tree_grpo_27b_mix4_hr24k_u6c (2026-07-20T23:08Z, 0 rows).
• Status: finished, 2026-07-19 to 2026-07-21 (26 W&B history rows and 23 timing.jsonl rows total across the surviving runs). hr/hr_b (fresh start, caps 8192/12288): the first process crashed pre-rollout, its relaunch under W&B id hr_b logged steps 1-10 and was killed in round 11 on the user's cap-raise order. hr16k (resumed from hr/global_step_5, caps 8192/16384): logged steps 6-8, killed once the extraction contract was ready; its own rollout directory is LOST (section 5). hr32k (resumed from hr/global_step_5, caps 12288/32768, contract deployed): 0 steps, update_actor OOMed at the 45056-token packing after round 6. hr24k (resumed from hr/global_step_5, caps 12288/24576, util 0.3): 0 steps, killed mid-round 6 on approval for the utilization change. hr24k_u6 (same caps, util 0.6): logged steps 6-13, killed after the shared-seed bug was diagnosed. hr24k_u6b (resumed from hr24k_u6/global_step_10, the seed fix applied): logged steps 11-15, update_actor OOMed stochastically at step 16. hr24k_u6c (resumed from hr24k_u6b/global_step_15): 0 steps, killed in round 16 on the user's 2026-07-21 stop order (pivot to distillation 02); the lineage's latest trained checkpoint remains hr24k_u6b/global_step_15.

2. Directory tree

The root holds one subdirectory per run id, plus this front page. No file currently sits at the root's own top level besides what a publish adds (README.md, FILES.txt); every path below already exists under analysis/rl-training/09-multitask-humanrank/. No other question's registry cites a run from this repository (its consumer list is empty).

README.md                                          this front page
FILES.txt                                          every published file with its byte size, added by the publish program
20260719_045416_27b_mix4_hr/                       leaf tree_grpo_27b_mix4_hr; the one rollout directory shared by report rows "hr" (W&B id burned pre-rollout, 0 steps) and "hr_b" (steps 1-10, report section "hr_b: the 12288 arm, steps 1-10")
  round_1/ .. round_11/                             round 11 was in flight when the cap-raise kill order landed; round 10 was all-debug
  timing.jsonl                                      steps 1-10 (the arm's own logged training steps)
  viz_dashboard_<task>.png, viz_tree_<task>.png     one pair per task (aerial-cactus-identification, dogs-vs-cats-redux-kernels-edition, text-normalization-challenge-english-language, text-normalization-challenge-russian-language)
  [no directory here for leaf tree_grpo_27b_mix4_hr16k: its own rollout directory 20260719_160958_27b_m4hr16k no longer exists anywhere; section 5]
20260719_221809_27b_m4hr32k/                       leaf tree_grpo_27b_mix4_hr32k; report row "hr32k" ("cap 32768 with the extraction contract"); resumed from hr/global_step_5
  round_6/                                          the only round; update_actor OOMed at the 45056-token packing right after it, so no timing.jsonl and no own checkpoint
  viz_dashboard_<task>.png, viz_tree_<task>.png     all 4 tasks
20260720_020550_27b_m4hr24k/                       leaf tree_grpo_27b_mix4_hr24k; report row "hr24k" (first half of "the 36864 rung"); resumed from hr/global_step_5, util 0.3
  round_6/                                          killed mid-round on approval to raise gpu_memory_utilization to 0.6, cutting every task's generation short before it reached the full 16: 12 aerial-cactus-identification, 10 dogs-vs-cats-redux-kernels-edition, 4 russian-language and 0 english-language node records; no task has a summary.json, and the 0-record english-language task also has no transcript.md or tree.txt (section 2 closing note); no step completed, no timing.jsonl
  [no viz_dashboard_<task>.png or viz_tree_<task>.png under this run at all: the mid-round kill landed before the harness rendered any of them, unlike every other training-arm run directory below]
20260720_041422_27b_m4hr24ku6/                     leaf tree_grpo_27b_mix4_hr24k_u6; report row "hr24k_u6"; resumed from hr/global_step_5, util 0.6
  round_6/ .. round_14/                              round 14 rolled out but the seed-bug diagnosis and the user's approval killed the arm before its update finished
  timing.jsonl                                      steps 6-13
  viz_dashboard_<task>.png, viz_tree_<task>.png     all 4 tasks
20260720_160605_27b_m4hr24ku6b/                    leaf tree_grpo_27b_mix4_hr24k_u6b; report row "hr24k_u6b" ("the seed fix"); resumed from hr24k_u6/global_step_10
  round_11/ .. round_16/                             round 16 rolled out fully (64 node records across all 4 tasks) but its step-16 update_actor died in the same stochastic 36864 OOM, so it has no timing row
  timing.jsonl                                      steps 11-15
  viz_dashboard_<task>.png, viz_tree_<task>.png     all 4 tasks
20260720_222757_27b_m4hr24ku6c/                    leaf tree_grpo_27b_mix4_hr24k_u6c; report row "hr24k_u6c" ("the end of the series"); resumed from hr24k_u6b/global_step_15
  round_16/                                          the only round (60 node records: cactus 16, dogscats 16, en 14, ru 14, per the report); killed on the user's 2026-07-21 stop order before this round's update ran
  viz_dashboard_<task>.png, viz_tree_<task>.png     only aerial-cactus-identification and dogs-vs-cats-redux-kernels-edition — the two text-normalization visualizations were never rendered before the kill
probe_bigresults_denoising/                        pre-launch wiring check (2026-07-19, report section "Verification"): the harness's 64 MiB node-results cap tested end to end, denoising-dirty-documents, model gpt-5.6-luna (not the trained policy)
  round_0/                                          2 nodes, both valid
  viz_dashboard_denoising-dirty-documents.png, viz_tree_denoising-dirty-documents.png
probe_bigresults_textnorm_en/                      the same 64 MiB cap check, text-normalization-challenge-english-language, gpt-5.6-luna
  round_0/                                          2 nodes, both valid
  viz_dashboard_text-normalization-challenge-english-language.png, viz_tree_text-normalization-challenge-english-language.png
smoke_hr_denoising/                                pre-launch wiring check of the humanrank reward, denoising-dirty-documents, gpt-5.6-luna
  round_0/                                          2 nodes, both crashed (the sentinel path only, per the report)
  viz_dashboard_denoising-dirty-documents.png, viz_tree_denoising-dirty-documents.png
smoke_hr_textnorm/                                 the same humanrank wiring check, text-normalization-challenge-english-language, gpt-5.6-luna
  round_0/                                          2 nodes, both valid (rewards 0.793 and 0.851 per the report)
  viz_dashboard_text-normalization-challenge-english-language.png, viz_tree_text-normalization-challenge-english-language.png

The four pre-launch directories (probe_bigresults_denoising, probe_bigresults_textnorm_en, smoke_hr_denoising, smoke_hr_textnorm) are not leaves of their own: they belong to leaf tree_grpo_27b_mix4_hr's card and the report's "Verification" paragraph, and settle no criterion. A completed round_N/ directory holds four files per task attempted that round: <task>__<epoch-seconds>.jsonl (16 node records per task per round for a training-arm run, 2 for the four pre-launch checks; section 4), <task>__<epoch-seconds>__transcript.md (a human-readable rendering: node count, valid count, best score, then the tree), <task>__<epoch-seconds>__tree.txt (a compact one-line-per-node ASCII rendering of the same tree, ending with the group's valid count and best node id), and <task>__summary_<epoch-seconds>.json (the harness's own per-task summary of the round; section 4). This shape holds for every round except the last, in-flight round of a lineage that was cut short by that run's own kill or stop event (section 1): in 20260719_045416_27b_mix4_hr/round_11 (the cap-raise kill) aerial-cactus-identification stopped at 7 node records and text-normalization-challenge-english-language at 14, and neither of those two has a summary.json (dogs-vs-cats-redux-kernels-edition and the Russian-language task both reached the full 16 and do have one); in 20260720_041422_27b_m4hr24ku6/round_14 (the seed-bug kill) both text-normalization tasks stopped at 9 (English) and 8 (Russian) records with no summary.json, while cactus and dogscats reached 16 and have one; in 20260720_222757_27b_m4hr24ku6c/round_16 (the 2026-07-21 stop order) both text-normalization tasks stopped at 14 records with no summary.json, while cactus and dogscats reached the full 16 and have one (matching the 60-node count already given for this run above); and in 20260720_020550_27b_m4hr24k/round_6 (the utilization-change kill) every task was cut short before reaching 16 — 12, 10, 4 and 0 node records for cactus, dogscats, Russian- and English-language respectively — no task in that round has a summary.json, and the English-language task, at 0 records, also has no transcript.md or tree.txt, consistent with the harness only writing those two rendered files once a task's round holds at least one node. No run directory holds a checkpoint, a verl file logger, a W&B client directory, or a console log: the lineage's trained LoRA weights and verl's own checkpoint state live outside this root, at /data2/peijia/ml-agent-checkpoints/tree-grpo/<run>/ (only tree_grpo_27b_mix4_hr, tree_grpo_27b_mix4_hr24k_u6 and tree_grpo_27b_mix4_hr24k_u6b have one there); the step-level training metrics (reward_node_mean, kl_loss, ppl_ratio, ...) live only in the W&B history above and its export copy under personal/experiments/rl-training/09-multitask-humanrank/results/wandb_exports/<run>/ in the personal repository; and each arm's console log is analysis/logs/tree_grpo_27b_mix4_<arm>.log, a project-wide directory outside this per-question root.

3. How to read each kind of file

• Small text, JSON and JSONL files (tree.txt, transcript.md, summary.json, timing.jsonl): https://huggingface.co/datasets/t2ance/mat-09-multitask-humanrank/raw/main/<path>, for example raw/main/20260719_045416_27b_mix4_hr/timing.jsonl or raw/main/20260719_045416_27b_mix4_hr/round_1/aerial-cactus-identification__1784437355__tree.txt.
• The per-task node-record JSONL files (2 to 12 MB each — they carry the full prompt, the raw completion, the token ids and the per-token logprobs of all 16 siblings) and every PNG image: resolve/main/<path> (raw/ returns only the LFS pointer for a tracked file), for example resolve/main/20260719_045416_27b_mix4_hr/round_1/aerial-cactus-identification__1784437355.jsonl or resolve/main/20260719_045416_27b_mix4_hr/viz_dashboard_aerial-cactus-identification.png.
• No adapter is published in this repository: every arm's summary.json carries an empty adapter_path (section 4) because the harness serves the in-memory trained weights to its rollout pods directly, and the lineage's actual checkpoints live outside this root (section 2, section 5).
• The reward and its invariant, per the report's Preliminary: reward in a node record is the humanrank percentile in [0, 1] for a valid node (harness/reward.py, section 1), and the fixed sentinels for every other node: -0.5 for crashed/timeout/no_code/no_submission/invalid_submission, -0.2 for an idle node (exc_type: CpuIdleError), -1.0 for an infrastructure failure; the report's step-1 health gate checks exactly this — every task has at least one valid node and every valid node's reward is inside [0, 1]. status: no_code marks a completion that ended inside an unclosed code fence under the 2026-07-20 extraction contract (harness/prompt.py, harness/search.py): code is empty and response_token_ids reaches the arm's own response cap, which is what the report's resp_clip_rate and "clipped at the cap" node counts (section 4 tables) check len(response_token_ids) against — the caps by arm (prompt/response/total, from each run's own config.json and launch script): hr/hr_b 8192/12288/20480, hr16k 8192/16384/24576, hr32k 12288/32768/45056, hr24k/hr24k_u6/hr24k_u6b/hr24k_u6c 12288/24576/36864.
• The other per-round instruments the report reads (groups_with_signal_rate, draft_rate, valid_reward_mean, draft_reward_mean, reward_best, reward_node_mean, no_code_rate) are all recomputable from a round's four node-record JSONL files: group the 16-sibling sibling_group values by task to test whether a group's rewards are all identical (the necessary condition the report's "signal density" paragraph checks); mode gives the draft/improve/debug split; reward and status give the rest. The policy-side training metrics the report also tabulates (kl_loss, ppl_ratio, actor/grad_norm, entropy) are not derivable from anything in this root — they exist only in the W&B history at the URLs in section 1 and in its export copy cited in section 5.

4. Row fields

round_N/<task>__<epoch-seconds>.jsonl, one row per candidate node of that task's round-N expansion (16 siblings, 2 for the four pre-launch checks): id, parent_id (null for a draft node); depth; mode (draft answers the task prompt from scratch, improve extends a valid parent, debug repairs a crashed one, per the report's Background); round; step (verl's global training-step counter at rollout time — recorded as 0 in every record inspected, so the file's own round number is the real index, not this field); sibling_group (the id shared by all siblings of one expansion — GRPO's advantage-contrast unit); exp_id (the MLE-bench competition slug); source (e.g. grade); status (ok, crashed, timeout, no_code, no_submission, invalid_submission, ...); is_buggy; exc_type (the raised Python exception class, e.g. NameError, ValueError, TimeoutError, CpuIdleError); grade_error; term_tail (the tail of the executed script's stdout/stderr); returncode; test_metric (the competition's own raw metric, before the humanrank transform); oriented (the same metric oriented so higher is always better); parent_oriented, parent_P, parent_P_evidence, improve_delta (populated on improve/debug nodes); group_best (whether this node is its group's best); medal; n_children; reward (the value GRPO actually trains on: the humanrank percentile or a crash/idle/infra sentinel, section 3); decl_deps_ignored, decl_malformed, decl_miss_reason; attempt; plan (the policy's stated plan text); where (the k8s pod/job id); resources (the pod's requested resources, e.g. {gpu:1,mem:16}); exec_time; exec_timings (a dict: pending, stage, exec, retrieve); cpu_util_mean, cpu_util_peak, gpu_util, gpu_util_peak, gpu_mem_used_gi, mem_peak_gi (null on CPU-pod tasks); gpu_name, kube_node; prompt_messages (the full chat prompt sent to the policy, including the system prompt's hard requirements and the extraction contract's rule); raw_completion (the policy's full raw output); thinking; code (the extracted Python script; empty when status is no_code); response_token_ids, response_logprobs (the completion's token ids and per-token logprobs — len(response_token_ids) is the field section 3's cap check uses).

round_N/<task>__summary_<epoch-seconds>.json, one row per task per round: exp_id; round; model (the serving model id — the trained policy for the eight training arms, gpt-5.6-luna for the four pre-launch checks); exec_backend (k8s); adapter_path (empty in every record inspected, section 3); nodes; valid; buffer_nodes; expansions; phase (train); best_test_metric; best_node_id; honest_test_metric (null in every record inspected); rollout_file (the harness's own internal relative path, e.g. data/rollouts/<run id>/round_N/<task>__<ts>.jsonl — not the path under this repository); workspace (the harness's internal sandbox id); gen_llm_seconds; gen_exec_seconds.

timing.jsonl (present only in 20260719_045416_27b_mix4_hr, 20260720_041422_27b_m4hr24ku6, 20260720_160605_27b_m4hr24ku6b — the three runs whose lineage completed at least one training step), one row per completed step: step; total_s; phases_s (a dict: update_weights, gen, old_log_prob, ref, adv, update_actor — the wall-clock seconds of each phase).

<task>__<epoch-seconds>__tree.txt and <task>__<epoch-seconds>__transcript.md are rendered text, not row files (section 3): the tree file is one line per node ([id] mode status test=... sib_grp=...) followed by a nodes=/sibling_groups= summary and the best valid node; the transcript adds the round's node/valid/best-score header and a ## Tree section with per-node plan/code/result detail.

5. Exclusions and inputs

The rule's fixed exclusion list, as it applies here: optimizer state, RNG state and other extra state of a training step, and full-model shards of a LoRA step — none of these is present under this root at all, because this harness's rollout output (round_N/, timing.jsonl, the viz PNGs) never includes them; the lineage's actual optimizer/RNG state and fp32 shards live in verl's own checkpoint directories outside this root, at /data2/peijia/ml-agent-checkpoints/tree-grpo/<run>/. The critic or value model: none exists — every arm uses algorithm.adv_estimator=grpo_valid_only (mean-only advantage, no critic). Resume pointers such as a latest-step file: not present under this root either; the resume path from one arm to the next (e.g. hr16k's trainer.resume_from_path=/data2/peijia/ml-agent-checkpoints/tree-grpo/tree_grpo_27b_mix4_hr/global_step_5) is read from the launch script, outside this root. An intermediate full-model dump next to a saved adapter: not present, same reason. Beyond the fixed list: this question's data root, as laid out, holds only tree-search rollout evidence — no run directory here contains the lineage's trained LoRA weights at all (they are the checkpoints named above, under /data2/peijia/ml-agent-checkpoints/tree-grpo/tree_grpo_27b_mix4_hr, ..._hr24k_u6 and ..._hr24k_u6b, one directory per kept step 5/10/15), so nothing was excluded from a "checkpoints/" role here — there is no such role in this question's actual layout, unlike the rule's usual template.

Inputs from elsewhere (not copied here):

• The base model: Qwen3_5ForConditionalGeneration at /data2/peijia/models/qwen3_5_27b on the workspace host, per actor_rollout_ref.model.path=${MODEL_PATH} in every arm's launch script (MODEL_PATH=/data2/peijia/models/qwen3_5_27b); the report's own text calls it "Qwen3.6-27B" (abstract; "Held constant" section) — both facts are reported as found, without resolving the naming difference. Every arm trains a LoRA adapter (r32/alpha 32, target_modules=all-linear, exclude_modules='.*visual.*') from scratch on this base; no pretrained adapter is an input to the hr leaf, and every later leaf resumes from a checkpoint produced inside this same question's own lineage (not from another question), at the paths named above.
• The tasks: the trained mix is four MLE-bench competitions — aerial-cactus-identification, dogs-vs-cats-redux-kernels-edition, text-normalization-challenge-english-language, text-normalization-challenge-russian-language — selected by the pre-audit over a fifth candidate, denoising-dirty-documents, which the audit rule swapped out (validity-only headroom: d_qual 0.026, spread 0.009) but which the four pre-launch checks still exercise; all five are the harness's packed task caches and the MLE-bench private answers and leaderboards (leaderboard.csv, read by the humanrank reward), benchmark data that stays outside this repository.
• The pre-audit's comparison archive: the gpt-5.6-luna search probes of 2026-07-18/19 (the same model that ran the four pre-launch checks in this root) and this question's own archived policy rollouts, both summarized in personal/experiments/rl-training/09-multitask-humanrank/results/preaudit_multitask_20260719_final.txt in t2ance/ml-agent-training-personal; the raw archived rollouts themselves are not this question's own runs and are not copied here.
• The report, cards, launch scripts, W&B export copies, and the one OOM log copy: all live under personal/experiments/rl-training/09-multitask-humanrank/ (cards, configs/, results/) in t2ance/ml-agent-training-personal; cited by that path throughout this README, never copied into this dataset.
• The LOST input, 20260719_160958_27b_m4hr16k (the hr16k arm's own rollout directory, rounds 6-9): a dangling symlink into a since-deleted personal/training_verl/ tree. The hr16k card's own 2026-07-19 17:50 entry explains why: the tmux launch inherited the wrong shell working directory, so the harness first wrote rollouts to personal/training_verl/runs/qwen3_5_27b/data/rollouts/ (a relative path) rather than the canonical analysis/rollouts/; the run was symlinked into the canonical location as a fix, but the personal/training_verl/ tree it pointed into was later deleted, and no copy survives under this repository, in analysis/rollouts/, or in any other question's published repository. Only the report's own numbers survive it: the round-6 breakdown in the report's "hr16k: the cap raise alone" section (31 of 64 nodes clipped at 16384, 18 NameError crashes, 0 rescued text-normalization solutions between 12288 and 16384) and the W&B history at tree_grpo_27b_mix4_hr16k (3 rows) with its export copy at personal/experiments/rl-training/09-multitask-humanrank/results/wandb_exports/tree_grpo_27b_mix4_hr16k/ in the personal repository.