拓展阅读 / NanoJev：小型决策模型

知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

Mixed expert SFT: completed game evaluation

One shared NanoJev checkpoint now reaches 27/128 Predict Position test successes, up from 11/128 before this training round. Snake improves from 6/8 to 8/8, Basic remains 128/128, and Maze remains 4/10. The matched Jev run scores 11/128 on Predict Position, 56/128 on Basic, 8/8 on Snake and 7/10 on Maze. This round improves shooting and Snake; Maze remains the weaker task in the comparison.

Test results

Each cell is successful episodes / evaluated episodes.

Predict Position success rises from 8.59% to 21.09%. Against Jev, the selected model wins 25 paired cases and loses 9; the two-sided exact McNemar p-value is 0.00904 before multiple-comparison adjustment. The per-model Wilson 95% intervals are 14.92–28.95% for NanoJev and 4.87–14.73% for Jev.

OOD results

Predict Position OOD changes the action cadence from four to eight physical ticks. NanoJev records 10/128 successes versus Jev's 8/128, with 10 paired wins, 8 losses and an exact p-value of 0.8145. Maze and OOD Snake remain below Jev in this cohort. These are the main gaps for the next round.

Data and selected model

• Collected 896 expert episodes / 17,498 decisions with the released Sonic Doom visual policy.

• Prepared 11,173 Predict Position questions across five splits, including 6,788 training questions from 512 training episodes. The other 384 expert episodes belong to dev, calibration, test and OOD.

• Preserved 7,587 Maze, Snake and Basic rows, their exact bytes and original splits. Each hard/soft dataset contains 18,760 rows in total.

• Trained four matched arms: hard/soft action targets × backbone learning rates 1e-5/2e-5. All completed 600 optimizer updates with training seed 17.

• Selected hard_lr1e5, checkpoint step 400, using development data only. Its weighted development success is 80.31%, versus initialization's 69.64%.

• Retained the shared Qwen3-0.6B backbone and dynamic Choice head. All four tasks use the same checkpoint and complete-question cross entropy.

The 600-step sampler draws 2,400 Predict Position questions, covering 2,021 distinct questions. The selected step-400 checkpoint uses the corresponding shorter prefix of that schedule. This is a fixed-update training comparison.

Evaluation protocol and verification

All four primary systems play the same 548 frozen cases: 274 test and 274 OOD. They receive the same structured observation interface, candidate actions, environment rules and seeded epsilon-greedy controller (epsilon 0.1, seed 17). The weighted score assigns Maze and Snake 1/3 each, Basic and Predict Position 1/6 each. Navigation uses its existing shared code planner. Shooting success requires an actual kill before the episode deadline.

Untuned Qwen uses the original Qwen3-0.6B weights, official chat template and native vocabulary head, with probabilities restricted to the offered A–D answer tokens. It receives no project training.

All 2,192 primary episodes pass independent simulator replay with zero mismatches. The frozen case registry, actual model identities, seeded action sampling and recorded physical counters also pass the result validator. The machine-readable report contains per-task intervals, paired case IDs, physical metrics and source hashes.

The visual expert is a separate reference: RGB frames, recurrent state and pure greedy actions. It succeeds on 125/128 PP test and 61/128 PP OOD cases. The 64-case development controller diagnostic records 62/64 with pure greedy actions and 44/64 with epsilon 0.1. The primary comparison above keeps its predeclared common controller.

Artifacts and reproduction

Local experiment artifacts are organized as follows:

All four complete model bundles also have verified persistent server backups under /data/rwang/nanojev_sonic_sft_20260920/checkpoint_archives/. The selected bundle's 15 files pass local SHA256 verification. The current model and complete mixed dataset are available through the public Hugging Face repositories NanoJev and NanoJev-Data, version unified-games-v1. See the unified release contents and verification.

Selected weights SHA256: f68c47d66998231b86b7e91b4ed5e82ae23acf104c8b7cd6d165c3ac7b7ffe1b.

See collection and training commands. To regenerate the full report in a fresh output directory:

python scripts/summarize_sonic_supervision.py \
  --cases runs/sonic_unified_sft_v1/test_cases.jsonl \
  --experiment runs/sonic_unified_sft_v1/experiment \
  --jev runs/sonic_unified_sft_v1/jev_parallel_recovery_v2/episodes.jsonl \
  --native runs/sonic_unified_sft_v1/native_test.jsonl \
  --expert runs/sonic_unified_sft_v1/expert/episodes.jsonl \
  --expert-protocol configs/sonic_predict_supervision_v1.json \
  --output runs/sonic_unified_sft_v1/report_reproduction

快照未包含部分工程附件或配图；请通过上游链接或原始材料查看，收录范围见来源清单。