File size: 5,482 Bytes
e4119e6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 | {
"schema_version": 1,
"model_repository": "baseten/caesar-cipher-adam-qwen38_27b-expert",
"environment": "caesar-cipher",
"environment_display_name": "Caesar Cipher",
"base_model": {
"id": "Qwen/Qwen3.8-27B",
"revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
"tokenizer_revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
"license": "apache-2.0",
"license_source": "https://huggingface.co/Qwen/Qwen3.8-27B/blob/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0/README.md"
},
"upstream_environment": {
"repository": "https://github.com/open-thought/reasoning-gym",
"revision": "49b07130b3fcd12f2d064bba7c43869543a0e7e7",
"path": "reasoning_gym/algorithmic/caesar_cipher.py",
"url": "https://github.com/open-thought/reasoning-gym/blob/49b07130b3fcd12f2d064bba7c43869543a0e7e7/reasoning_gym/algorithmic/caesar_cipher.py"
},
"training": {
"mode": "full-weight",
"algorithm": "grpo",
"optimizer": "adam",
"learning_rate": 1e-06,
"lr_decay_style": "constant",
"adam_betas": [
0.9,
0.98
],
"weight_decay": 0.1,
"clip_epsilon": 0.2,
"clip_epsilon_high": 0.28,
"kl_coefficient": 0.0,
"entropy_coefficient": 0.0,
"completed_updates": 100,
"training_tasks": 114,
"validation_tasks": 50,
"questions_per_update": 8,
"completions_per_question": 8,
"consumed_training_completions": 6400,
"schedule_seed": 20261019,
"training_dtype": "bfloat16",
"tensor_parallel_size": 8,
"async_batching": "ordered",
"max_completed_update_lag": 1
},
"generation": {
"enable_thinking": true,
"reasoning_effort": "medium",
"temperature": 1.0,
"top_p": 0.95,
"top_k": 20,
"max_output_tokens": 16384,
"context_tokens": 32768
},
"evaluation": {
"metric": "mean binary correctness across all sampled completions, including truncation failures",
"aggregation": "50 held-out tasks times 8 completions; not best-of-8 or pass-at-8",
"seeds": [
17,
43,
101,
137,
211,
307,
401,
503
],
"interval_updates": 10,
"dedicated_evaluation_role": true,
"baseline": {
"completed_updates": 0,
"accuracy": 0.39,
"correct_completions": 156,
"scored_completions": 400,
"held_out_tasks": 50,
"truncated_completions": 1,
"truncation_rate": 0.0025,
"observed_at": "2026-10-19T08:19:13.389311+00:00"
},
"final": {
"completed_updates": 100,
"accuracy": 0.9525,
"correct_completions": 381,
"scored_completions": 400,
"held_out_tasks": 50,
"truncated_completions": 13,
"truncation_rate": 0.0325,
"observed_at": "2026-10-19T13:25:51.973586+00:00"
},
"accuracy_gain_percentage_points": 56.25,
"checkpoint_selection": "final optimizer update; no best-checkpoint selection",
"limitations": "Same-environment held-out case evaluation. Not evidence of transfer to other environments, new task families, or a separately held-out final confirmation set. The same validation tasks were evaluated throughout training.",
"finite_corpus_note": "114 training plaintexts from the finite approved corpus, with 50 separate validation plaintexts and no train/validation plaintext overlap"
},
"checkpoint_provenance": {
"training_job_id": "q08d01w",
"training_project_id": "q41er1q",
"attempt": "run-06448b429f21",
"experiment": "miles-qwen3.8-27b-06448b429f21",
"training_source_commit": "a1fdca368c26ff0bfc08f498524b72591f610649",
"source_dirty": false,
"configuration_sha256": "bc83c25b9ee2f565235b7121bc165d9f4f875bcdea7db2462f2def97e20f47ce",
"selection_id": "3fb99bb437675fac893962e4a2e6db07e53d1ec9c62cdc2cf29c257637c3d66f",
"runtime_sha256": "77446fce2711b2302bc99b3ba1577f172e7d5cb8eb262cbeb356c64848767162",
"split_manifest_sha256": "861eca0de5c81c4a16a3e4eb3751aaa22ee26830443342c985922ac7815914ce",
"grading_sha256": "a77eff68098e44e75abb5477d23f13577435ca10d2fc8ff55cf7a2ac2fda2434",
"reward_contract": "rg-caesar-cipher-final-answer-v2-completed-only",
"inference_checkpoint_id": "miles-qwen3.8-27b-06448b429f21/run-06448b429f21/training/eval-snapshots/step_99",
"inference_checkpoint_size_bytes": 53815103533,
"inference_archive_status": "COMPLETE",
"native_completion_timestamp": "2026-10-19T13:26:09.282554+00:00",
"native_final_evaluation_receipt_sha256": "78a09ee7d4ae55ebf4a2033f403d251cc5c7aa182cb843b373aa139653ebe25b"
},
"runtime_versions": {
"miles": "a22fd824191c64ba7c5d6c531f125406e4ce9cc6",
"megatron": "fd15ee20a4f03b03680529baf4b8f7eeed64df1d",
"sglang": "b84bec5761a6cdc8b1df86b9b89ead80a7a9bb02",
"transformers": "5.12.1"
},
"reporting": {
"wandb_url": "https://wandb.ai/baseten-training/qwen38-reasoning-gym-adam100/runs/lo5clpo9",
"native_final_metrics_verified": true,
"final_point_present_in_wandb": false,
"last_reported_validation_update": 90,
"last_reported_validation_accuracy": 0.95,
"issue": "The final evaluation point is missing from W&B. Existing-run updates are permission-blocked; this release uses verified native final evaluation receipts and archived rollouts."
},
"release_scope": "Final inference weights and required inference assets with this sanitized model metadata. Raw task packages, raw response traces, credentials, machine paths, and optimizer state are not part of this release."
}
|