t2ance/mat-08-ropd-round-shape

08. Sixteen responses per state with a shared Luna teacher

1. Question and links

Read this first. The GitHub reading copy is for case studies: it retains full text, code, scores, execution results, and W&B evidence. It omits per-token probability arrays, input/output token ID arrays, tokenizer.json, and saved training steps. The Hugging Face archive retains the original selected records and saved adapters. Each copy has its own FILES.txt and SHA256SUMS. Both copies have been uploaded and their file listings checked.

The question is whether the selected sixteen-response recipe improves execution validity and held-out task performance. The student is Qwen3.5-9B with an SFT adapter. GPT-5.6 Luna at low effort supplies the reference, rubric, and verifier roles. Each round has five task states and sixteen responses per state.

• Report PDF and source.

• Run card and configuration, also copied under run_record/.

• Formal W&B run. Four client directories belong to this same run across launches and recoveries.

• Issue: none assigned. Review comments can identify the report section and exact source file.

• Snapshot status: training and independent evaluation complete. Launch A is global step 1. The eight continuation rounds B1–B8 are global steps 2–9. The archive contains 720 unique training candidates. The primary window has 80 valid candidates out of 240, compared with 36 out of 240 in the baseline window. Both validity windows pass their fixed threshold. The returned evaluation supports the registered exploratory rule; see ropd-train-wide/evaluation_final_verdict.json and the report for the valid-attempt result and final-solution limitation.

• W&B checks: nine history rows and 1,329 shared values agree with native metrics. The system-history API returned 3,148 rows for samples=100000, while its server event counter reported 1,930. Both counts and raw records are retained. Neither count is claimed as the number of unique physical measurements.

• Known gaps: saved task-state directories are available for steps 1–6, 8, and 9. The original step-7 full checkpoint was pruned before its task-state directory was returned. Its adapter, candidate records, pending states, transitions, and checkpoint verification remain. The initial SFT adapter belongs to question 03 and has no verified public archive link in this snapshot. These gaps limit standalone reproduction.

• Publication status: published to the original GitHub reading-copy repository and the Hugging Face archive. The 4 GiB limit remains in force. Per-copy indexes and checksums describe each representation; archive/snapshot.json records the verification.

• Retained diagnostic W&B records: wb-heartbeat-20260908t181412z-default, wb-heartbeat-20260908t181412z-slowstats, wb-installed-20260908t2032, wb-offset-20260908t193230, wb-offset-fixed-20260908t193500, wb-runtime-20260908t193150-long, wb-runtime-20260908t193150-short. These are infrastructure checks, not additional formal training runs.

2. Directory tree

Same-shaped rounds and saved steps are folded. FILES.txt lists every published file and its byte size.

08-ropd-round-shape/
├── README.md
├── FILES.txt                         # File index for the copy being read
├── SHA256SUMS                        # Content hashes for the copy being read
├── schema.json                       # Row keys and observed types
├── archive/                          # Export and table-generation scripts, snapshot metadata
├── report/                           # Report source and PDF
├── results/                          # Teacher prechecks and report figures
├── run_record/                       # Card, configs, and experiment checks
├── tables/
│   ├── training_metrics.csv          # All native fields, one row per completed step
│   ├── training_metrics.json         # Same rows as JSON
│   ├── task_validity.csv             # All 45 task-round rows and denominators
│   ├── reward_branches.csv           # Counts by stored rule, validity, and penalty
│   ├── decision.json                 # Fixed-window arithmetic
│   ├── training_overview.*           # PNG and PDF
│   └── gpu_system_metrics.*          # PNG and PDF
└── ropd-train-wide/
    ├── config/                       # Effective trainer settings
    ├── logs/                         # Returned process and resource logs
    ├── verl_metrics/distillation/    # Native metric files from all launch segments
    ├── wandb/wandb/run-*/            # Four client directories, same formal run ID
    ├── wandb_export/ropd-train-wide-rounds/
    │   ├── run.json                  # State, export time, server counters
    │   ├── config.json
    │   ├── summary.json
    │   ├── history.csv               # Unique verified training steps
    │   ├── history_scan_raw.*        # Original scan in JSON and CSV
    │   ├── history_api.json          # Separate history API response
    │   ├── system_metrics.*          # Returned system records in JSON and CSV
    │   └── reconciliation.json       # Native/W&B comparison and count limits
    ├── rollout/ropd-train-wide/
    │   ├── round_<N>/                # Nine completed rounds, N=1 through 9
    │   ├── round_5.attempt1/          # Interrupted attempt retained
    │   └── teacher_spend.jsonl
    ├── teacher_calls/worker-*/       # Teacher requests, responses, and usage
    ├── recovery/                     # Memory, replay, API, and W&B diagnostics
    ├── checkpoints/global_step_<N>/  # Nine saved adapters, Hub-only
    ├── checkpoint7_verification.json
    ├── checkpoint9_verification.json
    ├── training_final_metrics.json
    ├── summary_rounds.json           # Final process segment only, steps 7–9
    └── ...                           # Preflight and other returned records

The empty local stages/ directory is not a completed evaluation result. The later evaluation must add eval_sft.json, eval_step_9.json, and all attempt records.

3. How to read each kind of file

Start with training_metrics.csv and decision.json. Select Raw or Download raw file for machine-readable content. The GitHub raw URL form is https://github.com/t2ance/mat-experiments/raw/refs/heads/main/08-ropd-round-shape/<path>. Repository access is required.

On the Hub, use https://huggingface.co/datasets/t2ance/mat-08-ropd-round-shape/resolve/main/<path> for every file type. Use resolve for large JSONL files too, so the response contains the data rather than a storage pointer.

File kind	Read method
JSON	Parse one object or array. Null is missing or unknown, not zero.
JSONL	Parse each nonempty line separately. The sibling file has one candidate per line. Prepared samples use their own schema.
CSV/TSV	Read the header. Nested CSV cells contain JSON strings. Empty cells preserve absent fields.
Native metrics	Each row has step and data. Combine rounds.before_*.jsonl with rounds.jsonl. The derived table retains all nine steps.
W&B	history.csv holds unique training steps; raw scans and system rows are separate. Native .wandb files are source evidence. No W&B upload is needed to read the CSV exports.
Code/logs	Inspect the recorded program and messages. Failed attempts remain present. A replay duration is not a complete original-round duration.
PNG/PDF	Open directly. Figures come from retained metric rows and can be rebuilt with archive/generate_tables.py.
Saved adapters	Download the exact step from the Hub and load with PEFT on the recorded Qwen3.5 base model.
SHA256SUMS	Run sha256sum -c SHA256SUMS after downloading the complete Hub selection. Saved-step files are absent from GitHub by design.

Example adapter selection:

from huggingface_hub import snapshot_download
from peft import PeftModel

root = snapshot_download(
    repo_id="t2ance/mat-08-ropd-round-shape",
    repo_type="dataset",
    allow_patterns=["ropd-train-wide/checkpoints/global_step_9/lora_adapter/*"],
)
# First construct the supported Qwen3.5 base model with the recorded runtime.
policy = PeftModel.from_pretrained(
    base_model,
    root + "/ropd-train-wide/checkpoints/global_step_9/lora_adapter",
)

The final adapter SHA256 is 28eee4763ed35df1a682539fac8227c3897f2fffbcc9ad74241a6bc6956b9252. The snippet shows file selection and loading, not a complete serving environment.

4. Row fields

schema.json lists every observed top-level key and type in the five main JSONL families, every native metric column, and every returned W&B system column. It also expands execution-result and policy-identity keys. Raw fields are not renamed.

Fields	Meaning
run_id, step, round, global_step, task, chain, unit	Run, outer-round, task, and chain identity.
Group IDs and sibling_index	Prompt-group and candidate identity, five groups of sixteen per round.
is_behavior, behavior_index, behavior_action	Candidate selected before execution to control the next state.
prompt_messages, prompt_sha256, response_*	Prompt identity, generated text, thinking, token IDs, and sampled log-probabilities.
code, code_sha256, exec_*	Program and execution placement and timing.
result, valid	Benchmark scores, outcome, errors, budget, timing, model size, and log paths.
parent_result, anchor_result_before, better_than_anchor_before, debug_streak	The candidate's own starting state and anchor comparison, not an SFT comparison.
ropd_q, group_mean_q, reward, reward_rule, grpo_reward_mean, advantage	Teacher scores and execution-conditioned rewards; advantage subtracts the group mean.
reference_texts, rubric, verify_batches, verdicts, verifier_label, verdict_row	Teacher artifacts and blind scoring labels.
teacher_calls, cost_usd, cost_by_role	Recorded calls and known API-equivalent usage cost. Unknown costs remain unknown.
State *_before, *_after, history_*, policy_before, served_policy_id	Selected state transitions and policy identity.
Scheduler capacity_*, active_jobs, waiting, p90_*, mem_frac_max, load1, reason	Capacity changes and their measured inputs.
ours_env/valid_rate	Valid candidates divided by all eighty candidates.
ours_env/candidate_anchor_improved and rate	Count and share meeting the comparison rule against their own starting anchors.
ours_env/behavior_advance_rate	Share of five selected behaviors that advance the state.
ours_env/outcome/*	Execution-outcome counts, including failures.
ours_gen/*	Generation lengths and truncation observations.
ours_ropd/*	Teacher shape, score/reward distributions, agreement, and cost. Fresh rubrics make q_mean unsuitable as a fixed quality test.
ours_time/*	Collection clocks. Teacher and execution overlap. Engine sleep is included in execution. Collection excludes later optimization and saving.
actor/*, rollout_corr/*, timing_s/*, perf/*	Native optimization, serving agreement, time, and performance measurements.
W&B _step, _timestamp, _runtime, _wandb	Tracker position, time, runtime, and metadata; not extra optimizer steps.
W&B system.*	Returned system observations. The GPU plot converts memoryAllocatedBytes to GiB by dividing by 2^30.

5. Exclusions and inputs

The original archive excludes optimizer state, RNG state, scheduler/extra state, data-loader state, resume pointers, critics, and full model shards reproduced by a saved adapter. This does not delete or modify the original cloud checkpoints. GitHub also excludes saved-step directories, adapter files, tokenizer.json, and numeric arrays named response_logprobs, response_token_ids, prompt_ids, response_ids, or input_ids. These exclusions apply to nested teacher records and recovery records too; full text, scores, execution results, and scalar token counts remain. The retained run records, teacher outputs, logs, and recovery evidence are included.

The initial SFT adapter is an input from question 03. Its recorded identity is sft_9b_checkpoint-184, LoRA rank 32, alpha 64, with recorded SHA256 3fc6c3d76712a64e005f80280b98ca21faacb18d6a7b0e06daef4f3af5f03958. It has no verified public input-archive URL in this snapshot and is not silently copied into question 08's results.

The base model is Qwen/Qwen3.5-9B as named by the saved adapter configuration. Base weights are not duplicated. The MLE-bench task caches and private scoring files are benchmark inputs and are not copied. Task names and execution budgets are recorded in the report and configuration.

The two held-out tasks are spooky-author-identification and dogs-vs-cats-redux-kernels-edition. The planned comparison has two policies, two chains per policy/task, and four attempts per chain. This training snapshot does not claim completed evaluation.

The teacher's references, rubrics, verdicts, token usage, and requests are retained outputs. API credentials and account files are not archive contents.