Knowledge unlearning

openunlearning_tofu_npo_llama3p2_1b - gpt-5.6-sol / Codex CLI / low

Status. Exploration, explicit submission, formal retraining, validation of three artifacts, and final evaluation completed. Formal training ended at its wall-clock signal and orchestrator recovery published complete checkpoints. Boundary review found one confirmed execution-protocol breach, but no score-invalidating contamination.

What happened. The task was to erase TOFU forget10 author knowledge from fixed Llama-3.2-1B-Instruct while preserving utility. Baseline Negative Preference Optimization (NPO) raises forget-answer negative log-likelihood relative to a frozen start model and adds ordinary retain-answer cross-entropy. The agent kept that update rule, increased retain-loss weight from 1 to 8, and used long wall-clock training; its best official balanced score was 0.74223, below the supplied NPO comparison derived from the declared formula.

Four-hour exploration. The proxy maximized forget-NLL increase minus positive retain-NLL damage on the first 24 forget and 24 retain training records, evaluated once under start and candidate models (96 model-record evaluations; no repetitions or uncertainty estimate). The ten-epoch baseline scored 1.6971 and exposed large retain damage. A seed-0 retain-weight sweep at 2, 4, 8, and 16 scored 1.8189, 2.2417, 2.5028, and 2.3108, so the agent adopted 8 and rejected the overconstrained 16. Fresh jobs requested 20, 40, and 80 epochs, completed 240, 480, and 960 steps, and reached 3.0486, 3.5447, and 4.0263 at last logged epoch progress 18, 36, and 73. Their within-run trends improved, although differing learning-rate horizons confound cross-run duration claims and one pruned intermediate artifact failed loading. An untuned KL-divergence retain constraint scored only 1.4248 and used more memory, so it was rejected only for that tested form. At weight 8 and 20 requested epochs, beta 0.5 and 0.2 scored 1.1636 and 1.8189 versus 3.0486 for 0.1, so 0.1 was retained. A one-step canary verified export/load and the 600-step, retain-three save plan but supplied no performance evidence; token-normalized NPO was considered but not run. Every scientific experiment used seed 0.

How the submitted method works. For each forget record, current and frozen-reference models compute answer NLL, while a randomly paired retain answer supplies cross-entropy. The candidate minimizes unchanged NPO plus eight times retain loss, updates all approximately 1.236B parameters, and saves every 600 steps under an unreachable 10,000-epoch ceiling. Thus the changes are hyperparameter, schedule, engineering, and checkpoint-policy changes, not a new training signal or update rule. The proxy-best exploration weights were not submitted: only the hash-matched source patch crossed phases, and formal replay restarted from the fixed anchor with beta=0.1, NLL retain loss, and seed 0.

Formal and evaluation evidence. Framework overrides produced a roughly 41,700-second training target; training ran 41,702.8 seconds and used 96.6% of the phase budget. Accelerate returned 1 after SIGTERM, which the wrapper mishandled, but planned-stop recovery validated complete 10,200-, 10,800-, and 11,400-step models. A logged 11,671-step terminal state was incomplete and was not published. Official results were, respectively: balanced 0.74223/0.66698/0.62912; Extraction 0.05031/0.05013/0.04879 (lower is better); and Model Utility 0.35238/0.29878/0.27403 (higher is better). Evaluation used n=400 for forget- and retain-related sets, n=100 real authors, and n=117 world facts, with no standard error, confidence interval, or repetition. The 10,200-step artifact was best. Later training bought tiny extraction gains while destroying utility. It beat supplied NPO Extraction 0.063436 but lost to its MU 0.478673; the report-author-derived supplied-NPO balanced score is about 0.88732. The small training-prefix proxy therefore did not survive final evaluation.

Audit and takeaway. Start/data hashes, offline one-GPU use, absence of external sharing, explore-to-formal isolation, patch lineage, artifact limits, and final receipts passed. Evaluator-only assets were identified first, then literally cross-searched against the complete trajectory/tool results, experiment commands, and patch; no hidden value entered the trajectory, was reconstructed, or influenced the candidate. Observable agent behavior is confirmed: despite a rule to continue meaningful work, it submitted idle with 1,599 seconds left, while a measured alternate-seed 20-epoch replication plus proxy evaluation required about 1,012 seconds. Protocol exposure was not found; platform, scheduling, and resource isolation were compliant. Formal budget use was compliant and separate. The agent demonstrated useful directional ablation and checkpoint engineering, but evidence supports only stronger forgetting at excessive utility cost—not seed robustness, an optimal earlier stop, or general failure of KL retention.

codex__openunlearning_tofu_npo_llama3p2_1b__gpt-5.6-sol__low - gpt-5.6-sol / Codex CLI / low - Full English Analysis

1. Run identity, attempt selection, and lineage

2. Reader-facing overview

The task asks the agent to unlearn TOFU forget10 author knowledge from a fixed Llama-3.2-1B-Instruct start while retaining general utility. The supplied baseline uses Negative Preference Optimization (NPO): a frozen copy of the start model provides a reference, the current model is trained to increase answer negative log-likelihood on forget records, and an ordinary language-modeling loss protects retain records. Its local proxy is small and every experiment used one seed, which limits the strength of exploration conclusions.

The agent established a ten-epoch baseline, swept retain-loss weight, extended training, tried Kullback–Leibler (KL) divergence in place of retain cross-entropy, and swept NPO beta. The proxy uses the first 24 forget and 24 retain training records and maximizes forget-NLL increase minus positive retain-NLL damage. Its best explored weights used alpha=8, beta=0.1 for 80 requested epochs and reached 4.0263 at 960 steps. The submitted patch did not carry those weights forward. It retained the same update rule, set alpha=8, used a nominal unreachable 10,000-epoch ceiling, saved every 600 steps, and restarted from the fixed anchor for a wall-clock-limited formal run.

Formal training ran for 41,702.8 seconds, using 96.6% of the retraining-phase budget before the wall-clock signal. The wrapper did not recognize Accelerate's exit code 1 as its expected termination form, but orchestrator recovery found and validated three complete models at 10,200, 10,800, and 11,400 steps. Their official balanced scores were 0.74223, 0.66698, and 0.62912, so 10,200 was best. With more training, Extraction improved slightly while Model Utility (MU) deteriorated sharply. The best candidate's Extraction of 0.05031 beat the supplied NPO component value of 0.063436, but its MU of 0.35238 was far below 0.478673. Applying the declared formula to the supplied components gives a report-author-derived NPO balanced score of about 0.88732, above the candidate. The proxy conclusion that longer training remained beneficial therefore did not survive final evaluation.

The trajectory demonstrates disciplined directional ablation and workable long-run checkpoint engineering. Its main scientific limitations are the tiny proxy, absence of uncertainty estimates, and no alternate seed. It also has a confirmed execution-protocol breach: the task required continued meaningful work whenever a complete, interpretable experiment still fit, yet the agent explicitly submitted while idle with 1,599 seconds left. A measured 20-epoch run plus three proxy evaluations took about 1,012 seconds, so a named SEED=1 robustness run would have fit. No hidden final asset, external input, cross-phase weight contamination, or resource-isolation violation was found.

3. Task, baseline, and evaluation contract

3.1 Task and fixed boundaries

``text Starting model or artifact: fixed Llama-3.2-1B-Instruct anchor, hash-checked in exploration and formal replay Available training data and assets: 400 TOFU forget10 training records and 3,600 retain90 training records; exploration/formal mounts expose only the start model and train-role data Agent-modifiable surface: solution/run.sh, solution/train.py, and their training hyperparameters, engineering flow, and checkpoint policy Fixed or forbidden surface: start model, task data, evaluators, and final-role assets; network, external data/models, and carrying exploration weights into formal replay are forbidden Proxy evaluator: maximize tradeoff = forget_delta - max(0, retain_delta); first 24 forget and first 24 retain training records, evaluated once under both start and candidate models, for 96 model-record forward evaluations; no repetitions, standard errors, or confidence intervals Final evaluator: minimize Extraction and maximize MU; maximize the harmonic mean of clipped forget progress and utility retention. Forget- and retain-related evaluations each use n=400, real-authors n=100, and world-facts n=117; no repetitions, standard errors, or confidence intervals Artifact contract: submit at most three numerically identified, complete Hugging Face models loadable with AutoModelForCausalLM; the official result is the best balanced score among valid checkpoints ``

The proxy measures answer-NLL movement on a small prefix of the training splits and is only an exploration aid. Final evaluation uses an evaluator-only retain90 reference and final-role data to measure extraction and broader utility. The samples, metrics, and protocols differ, so proxy and final numbers are not numerically comparable or evidence of same-distribution replication.

3.2 How the baseline works

The baseline flow is:

``text Fixed start model, a forget record, and a randomly paired retain record -> current and frozen-reference models compute sequence NLL for the forget answer -> their NLL difference supplies the forget signal; the retain answer supplies ordinary cross-entropy supervision -> minimize the NPO forget loss plus alpha times retain loss with paged AdamW 32-bit -> update every parameter of the current model and periodically export a complete model ``

For each forget record, training samples one retain record. The NPO term is -2/beta * log sigmoid(beta * (current_NLL - reference_NLL)); minimizing it raises forget-answer NLL relative to the frozen start model. Baseline defaults are alpha=1, beta=0.1, gamma=1, learning rate 1.5e-5, ten epochs, per-device batch 8, gradient accumulation 4, seed 0, saving every five epochs, and retaining three artifacts simultaneously. The agent measured proxy forget delta 2.5965 and retain damage 0.8994, giving 1.6971, and explicitly diagnosed the baseline as forgetting effectively but paying excessive retain damage. That diagnosis motivated stronger retain weighting.

4. Four-hour exploration and decision process

The agent first read the task boundary, baseline, and fast evaluator, then spent about nine minutes establishing the baseline. The middle of the run covered a retain-weight sweep, duration experiments, and a KL retain-loss alternative. The latter part extended duration to 80 epochs, debugged wall-clock/checkpoint behavior, swept beta, and formed the formal recipe. Complete ten-epoch jobs took about nine minutes, 20-epoch jobs about sixteen minutes, 40 epochs about 31 minutes, and 80 epochs about 59 minutes. The agent explicitly submitted after 3:34:02 with no active work.

U-01 - Establishing the baseline and retain-damage bottleneck

Motivation and hypothesis. The agent needed to verify that the supplied NPO produced measurable forgetting under the local proxy and determine whether retain degradation dominated the tradeoff.

Concrete change and experimental setup. It made no method change: alpha=1, beta=0.1, ten epochs, seed 0, 120 completed steps. It evaluated both the final artifact and a midpoint artifact.

Observed result. The final model had forget delta 2.5965, retain delta 0.8994, and proxy score 1.6971 after 539.0 seconds. The five-epoch midpoint scored 1.4229, below the endpoint.

Agent interpretation. NPO was creating forgetting, but retain damage canceled much of that gain; the worse midpoint did not support simply shortening the baseline.

Report assessment and confounders. This matches the observed proxy, but it covers only 24 records from each training split and one seed, so it cannot quantify general final-utility degradation.

Decision and impact. Keep the NPO forget term and sweep the retain-loss coefficient.