🔬Modal Research Exchange

{
  "base_model": "Qwen/Qwen3.5-4B",
  "baseline": "dimuon-lite (DiMuon with Polar Express), beta1=0.9, grad_clip_norm=0.0. Same model, rank, data order, seed, renderer, batch size, steps and LR grid. The best-LR run of each arm is compared.",
  "batch_size": 64,
  "budget": {
    "gpu_hours": 1.8,
    "max_wall_minutes": 120
  },
  "command": "python - <<'EOF'\n# Custom part: the TriMuon config exists only on the fork; the baseline uses tinker.types.DimuonOptimizerConfig() / DimuonParams.\nimport os, tinker\nfrom tinker_cookbook.recipes.chat_sl import chat_datasets  # renderer + no_robots builder\nsc = tinker.ServiceClient(base_url=os.environ[\"TINKER_BASE_URL\"])  # fork URL for the TriMuon arm\nARM = os.environ.get(\"ARM\", \"pe\"); LR = float(os.environ.get(\"LR\", \"1e-4\"))\nopt_cfg = tinker.types.DimuonOptimizerConfig() if ARM == \"pe\" else tinker.types.TrimuonOptimizerConfig()  # Trimuon* added by the fork\ntc = sc.create_lora_training_client(base_model=\"Qwen/Qwen3.5-4B\", rank=32, optimizer=opt_cfg)\ntrain, test = chat_datasets.no_robots_batches(tc.get_tokenizer(), batch_size=64, max_length=2048, seed=0)\nfor step, batch in zip(range(80), train):\n    tc.forward_backward(batch, loss_fn=\"cross_entropy\").result()\n    P = tinker.types.DimuonParams if ARM == \"pe\" else tinker.types.TrimuonParams\n    tc.optim_step(P(learning_rate=LR, beta1=0.9, grad_clip_norm=0.0)).result()\nprint(\"eval_nll\", sum(tc.forward(b, loss_fn=\"cross_entropy\").result().metrics[\"loss:mean\"] for b in test) / len(test))\ntc.save_weights_for_sampler(name=f\"rx-trimuon-{ARM}-{LR}\")\nEOF",
  "dataset": {
    "hf_path": "HuggingFaceH4/no_robots",
    "n_examples": 5120,
    "name": "no_robots",
    "split": "train (eval on test)"
  },
  "eval": {
    "benchmark": "no_robots test split",
    "metric": "mean per-token NLL on assistant tokens",
    "n_examples": 500
  },
  "hypothesis": "With a Spindle fork that adds TriMuon, LoRA SFT on Qwen3.5-4B reaches held-out no_robots loss at least 0.01 nats lower than dimuon-lite (Polar Express) at its best LR, after the same 80 steps.",
  "lora_rank": 32,
  "max_seq_len": 2048,
  "method": "Fork modal-projects/spindle and add TriMuon as an optimizer option. It keeps DiMuon's recovered momentum and preconditioning rules. It replaces only the Polar Express polar step with two outer Polar Switch passes, then four cubic Newton-Schulz refinements folded into the rank x rank multiplier. Use the FP16 polar core and TF32 disabled, as in aaron's benchmark. Deploy the fork as a throwaway copy. Train rank-32 LoRA on Qwen/Qwen3.5-4B with a short tinker SDK SFT loop on no_robots. The loop uses the cookbook renderer for the model, forward_backward with cross_entropy, then optim_step. The treatment arm runs on the fork with TriMuon. The baseline arm runs dimuon-lite on the same fork (or shared Spindle) with identical code. Each arm sweeps LR over {5e-5, 1e-4, 2e-4}: 6 runs of 80 steps each, same data order and seed. Each run evaluates held-out loss at step 80 and logs per-step optimizer wall time. Qwen3.5-4B is a scaled-down stand-in for the source's Qwen3-8B rank-32 workload. Polar Switch and AdamW arms are follow-up specs, not part of this decision.",
  "optimizer": {
    "lr": 0.0001,
    "name": "trimuon (fork)"
  },
  "steps": 80,
  "success_criterion": "Validated if best-LR TriMuon eval NLL is at least 0.01 nats below best-LR dimuon-lite eval NLL, and TriMuon's median optimizer step time is not higher. Failed if the gap is under 0.01 nats or reversed. A rerun of the best PE config with a second seed is optional, if budget allows, to check that the 0.01 margin exceeds seed noise.",
  "title": "TriMuon vs dimuon-lite (Polar Express) for LoRA SFT on no_robots",
  "training_type": "sft"
}

[
  {
    "arm": "pe-lr5e-05",
    "step": 80,
    "uri": "tinker://bf5f1b79ede219a7cf1dd36412f15530/weights/rx-trimuon-pe-lr5e-05"
  },
  {
    "arm": "pe-lr0.0001",
    "step": 80,
    "uri": "tinker://9a64c029dc6ce9030e42482021de9095/weights/rx-trimuon-pe-lr0.0001"
  },
  {
    "arm": "pe-lr0.0002",
    "step": 80,
    "uri": "tinker://d6e72fa4872b9b9bf6941e6db87cb3d3/weights/rx-trimuon-pe-lr0.0002"
  },
  {
    "arm": "trimuon-lr5e-05",
    "step": 80,
    "uri": "tinker://d5543577fb52d3f5b78a9ef486622d76/weights/rx-trimuon-trimuon-lr5e-05"
  },
  {
    "arm": "trimuon-lr0.0001",
    "step": 80,
    "uri": "tinker://71547e645250150e2d5c75d854d6db8e/weights/rx-trimuon-trimuon-lr0.0001"
  },
  {
    "arm": "trimuon-lr0.0002",
    "step": 80,
    "uri": "tinker://6b1a39c63d9977ef11e769b3a5c34991/weights/rx-trimuon-trimuon-lr0.0002"
  }
]

{
  "pe": {
    "best_eval_nll": 2.1021933133864747,
    "best_lr": 0.0002
  },
  "trimuon": {
    "best_eval_nll": 2.102713969608129,
    "best_lr": 0.0002
  }
}

base_eval_nll_token_mean

2.209

best_lr_pe

0.0002

best_lr_trimuon

0.0002

best_eval_nll_pe

2.102

best_eval_nll_trimuon

2.103

effect_nats

0.0005207

median_optim_step_seconds_pe_best

0.05992

median_optim_step_seconds_trimuon_best

0.04563

• checkpoint · pe-lr5e-05 · step 80 · tinker://bf5f1b79ede219a7cf1dd36412f15530/weights/rx-trimuon-pe-lr5e-05
• checkpoint · pe-lr0.0001 · step 80 · tinker://9a64c029dc6ce9030e42482021de9095/weights/rx-trimuon-pe-lr0.0001
• checkpoint · pe-lr0.0002 · step 80 · tinker://d6e72fa4872b9b9bf6941e6db87cb3d3/weights/rx-trimuon-pe-lr0.0002
• checkpoint · trimuon-lr5e-05 · step 80 · tinker://d5543577fb52d3f5b78a9ef486622d76/weights/rx-trimuon-trimuon-lr5e-05
• checkpoint · trimuon-lr0.0001 · step 80 · tinker://71547e645250150e2d5c75d854d6db8e/weights/rx-trimuon-trimuon-lr0.0001
• checkpoint · trimuon-lr0.0002 · step 80 · tinker://6b1a39c63d9977ef11e769b3a5c34991/weights/rx-trimuon-trimuon-lr0.0002
• effective_config · effective-config.json sha256 08bffcd8c7ab
• code_revision · code-revision.txt
• script · rxlib.py sha256 eece1ee4580b
• script · train_arm.py sha256 20566d1b5e51
• script · probe.py sha256 d3b88e7c49a4
• script · loss_eval.py sha256 d43aefd37b6d
• script · preflight_modal.py sha256 df60fdc09e94
• code_patch · spindle-fork.diff sha256 b619ea2b7bd6
• code_patch · miles-rx-only.diff sha256 501f3ad2efa3
• code_patch · miles-overlay-full.patch sha256 079bbf867c7b
• predictions · predictions.jsonl sha256 f0f0d6c336f0

1. Nov 4, 18:17 UTC · succeeded · done: pe, trimuon · On a Spindle fork that adds TriMuon (trimuon_fp16), 80-step rank-32 LoRA SFT of Qwen3.5-4B on no_robots gives best-LR (2e-4) held-out NLL 2.10271 for TriMuon vs 2.10219 for dimuon-lite (PE), so TriMu…
2. Nov 4, 18:12 UTC · running · done: pe, trimuon · All 6 runs (2 optimizers x 3 LRs, 80 steps, rank-32 LoRA on Qwen3.5-4B, no_robots) finished on fork spindle-fork-run_36f230d7bf. Best LR is 2e-4 for both: TriMuon reaches 2.10271 and dimuon-lite (PE)…
3. Nov 4, 17:44 UTC · running · Both arms run on fork spindle-fork-run_36f230d7bf (shared Spindle returns HTTP 400 for dimuon). Optimizer parity preflight passed exactly; base-model held-out NLL is 2.2094 nats/token; fixing determi…
4. Nov 4, 17:35 UTC · running · Shared Spindle rejects dimuon (HTTP 400 'optimizer family dimuon is not supported'), so both arms run on a fork (spindle-fork-run_36f230d7bf) that overlays Miles dimuon-lite (#17) + TriMuon (#27 @571…

adaptation_manifests

[{'arm': 'pe-lr0.0001', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_hash': '82475988ce285b1e01dda2f5003f245a2e7b62db4933d012245832c966e3149c'}, {'arm': 'pe-lr0.0002', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_hash': '82475988ce285b1e01dda2f5003f245a2e7b62db4933d012245832c966e3149c'}, {'arm': 'pe-lr5e-05', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_hash': '82475988ce285b1e01dda2f5003f245a2e7b62db4933d012245832c966e3149c'}, {'arm': 'trimuon-lr0.0001', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_hash': '82475988ce285b1e01dda2f5003f245a2e7b62db4933d012245832c966e3149c'}, {'arm': 'trimuon-lr0.0002', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_hash': '82475988ce285b1e01dda2f5003f245a2e7b62db4933d012245832c966e3149c'}, {'arm': 'trimuon-lr5e-05', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_hash': '82475988ce285b1e01dda2f5003f245a2e7b62db4933d012245832c966e3149c'}]

correctness_preflight

{'required': True, 'paths': [{'id': 'dimuon_lite_pe_baseline_through_fork', 'fixed_input': {'sha256': 'c393ff74b3c470c7c79e2838bad47715c9800515d9ac202d3cdd6f4139b7bc1f', 'description': 'seed-0 FP32 LoRA factors A[32,2560], B[9216,32]; two fixed gradient sets (seeds 1,2); fresh Dimuon state per probe; momentum 0.9, lr 1e-4, damping 1e-4, rank exponent 0.175 (preflight_modal.py)'}, 'baseline_revision': '228b59d90074e6cec48090e033202c5a4aff3ff3', 'reference_revision': '228b59d90074e6cec48090e033202c5a4aff3ff3', 'implementation_revision': '7aa6a401f06ecdb2827ebc04eb0c81a9293799e5', 'precision': 'FP32 masters, BF16 momentum, FP16 Polar Express core, TF32 disabled, matmul precision highest (H100)', 'normalization': 'direction Frobenius norm sqrt(rank) per block, update = lr*(rank/32)^0.175*direction', 'errors': {'baseline_update_step1': {'max_abs_error': 0.0, 'reference_max_abs': 8.878763765096664e-06, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0000000000000013}, 'baseline_update_step2': {'max_abs_error': 0.0, 'reference_max_abs': 8.814036846160889e-06, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0000000000000013}, 'reference_update_step2_vs_pr27_pe': {'max_abs_error': 0.0, 'reference_max_abs': 8.814036846160889e-06, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0000000000000013}}, 'lr_scaling': [{'learning_rate': 5e-05, 'update_norm': 0.00039999998332940976}, {'learning_rate': 0.0001, 'update_norm': 0.0007999999666588215}, {'learning_rate': 0.0002, 'update_norm': 0.0015999999333176456}]}, {'id': 'trimuon_fp16', 'fixed_input': {'sha256': 'c393ff74b3c470c7c79e2838bad47715c9800515d9ac202d3cdd6f4139b7bc1f', 'description': 'same fixture as above'}, 'baseline_revision': '228b59d90074e6cec48090e033202c5a4aff3ff3', 'reference_revision': '5713fc16bb104cd587ce2229eb2ec2d63acf7b1d', 'implementation_revision': '7aa6a401f06ecdb2827ebc04eb0c81a9293799e5', 'precision': 'FP32 masters, BF16 momentum, FP16 Polar Switch core with 4 refinements (Triton), TF32 disabled (H100)', 'normalization': 'same as baseline', 'errors': {'reference_update_step1': {'max_abs_error': 0.0, 'reference_max_abs': 8.666422218084335e-06, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0000000000000013}, 'reference_update_step2': {'max_abs_error': 0.0, 'reference_max_abs': 8.929520845413208e-06, 'relative_l2_error': 0.0, 'cosine_similarity': 1.000000000000002}}, 'lr_scaling': [{'learning_rate': 5e-05, 'update_norm': 0.0003999999697334397}, {'learning_rate': 0.0001, 'update_norm': 0.0007999999394668804}, {'learning_rate': 0.0002, 'update_norm': 0.0015999998789337647}]}], 'acceptance': 'prespecified before measuring: atol=0, rtol=0 (bitwise parity, same code), LR-scaling relative drift <= 1e-6', 'notes': 'v1 LR-scaling measured with FP32 params showed 1.9e-6 drift for trimuon from float32 (p_new-p_old) cancellation; fixed the instrument (FP64 masters for the difference only), tolerance unchanged; v2 drift ~4e-15. The spec had no explicit preflight plan; research_exchange.correctness validator was not available in the workspace.', 'treatment_active_check': {'max_abs_error': 8.121132850646973e-07, 'reference_max_abs': 8.814036846160889e-06, 'relative_l2_error': 0.0946766722261113, 'cosine_similarity': 0.9955181679355461}, 'polar_vs_svd_rel_l2': {'pe': 0.09879163297882344, 'trimuon_fp16': 0.0006948141318078318}}

correctness_validation

{'status': 'unverified', 'required': True, 'checks': [], 'reason': 'Correctness preflight was not recorded.', 'evidence_source': 'executor report; remote execution is not attested', 'applicability': 'unknown', 'reported_verdict': 'refuted'}

evaluation

{'version': 1, 'evaluators': {'e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee': {'code_sha256': 'd8ae7378f8e7835775de262c441ecdbb0b14fa90ae09742d53ae1f26c4ba5d92', 'prompt_template': "tinker_cookbook renderer 'qwen3_5' (model_info recommended for Qwen/Qwen3.5-4B), conversation_to_datum(messages, max_length=2048, TrainOnWhat.ALL_ASSISTANT_MESSAGES); no_robots test split rows 0-499", 'decoding': {'temperature': 0, 'max_tokens': 0, 'stop': [], 'note': 'teacher-forced forward(cross_entropy); no generation'}, 'parser': "loss_eval.py LossEval('token_mean'): sum NLL over tokens with weight>0 / count", 'items': {'manifest_ref': 'predictions.jsonl', 'n': 500}}}, 'scores': [{'arm': 'pe', 'metric': 'nll_token_mean', 'value': 2.1021933133864747, 'evaluator_id': 'e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee', 'seed': 0, 'predictions_ref': 'predictions.jsonl'}, {'arm': 'trimuon', 'metric': 'nll_token_mean', 'value': 2.102713969608129, 'evaluator_id': 'e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee', 'seed': 0, 'predictions_ref': 'predictions.jsonl'}, {'arm': 'base', 'metric': 'nll_token_mean', 'value': 2.209437855162829, 'evaluator_id': 'e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee', 'seed': 0, 'predictions_ref': 'probe_result.json'}], 'comparisons': [{'metric': 'nll_token_mean', 'arms': ['pe', 'trimuon']}]}

evaluator_validation

{'status': 'invalid', 'issues': ['Evaluator e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee: item IDs or a hashed item manifest are required.', 'Evaluator e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee: identity does not match provenance hash.'], 'comparisons': [{'metric': 'nll_token_mean', 'status': 'comparable', 'arms': ['base', 'pe', 'trimuon'], 'evaluator_ids': ['e2d2ec49b9c8eedc08952d1fa001a784424d3dd175eb361950141c6bd74474ee'], 'missing_arms': []}]}

feasibility

{'status': 'unknown', 'problems': ['Re-projection required: lora_rank changed or was not recorded', 'Re-projection required: dataset changed or was not recorded', 'Re-projection required: steps changed or was not recorded', 'Re-projection required: n_examples changed or was not recorded', 'Primary endpoint changed or was not recorded'], 'wall_seconds': None, 'gpu_hours': None, 'known_gpu_hours': 0.6602666666666666, 'phases': {'setup': {'wall_seconds': 900.0, 'planned_units': 1, 'planned_samples': 0.0, 'gpu_hours': None}, 'training': {'wall_seconds': 571.1999999999999, 'planned_units': 480, 'planned_samples': 30720.0, 'gpu_hours': 0.6346666666666666}, 'evaluation': {'wall_seconds': 23.04, 'planned_units': 3000, 'planned_samples': 3000.0, 'gpu_hours': 0.025599999999999998}, 'recovery': {'wall_seconds': 1200, 'gpu_hours': None}}, 'generation_counts': None, 'generation_applicability': {'check': 'generation_audit', 'status': 'unknown', 'enforce': False, 'reasons': ['spec has no experiment_features: generated_training_data'], 'features': {'generated_training_data': None}}, 'record': {'config': {'base_model': 'Qwen/Qwen3.5-4B', 'arms': ['pe', 'trimuon'], 'steps': 1, 'batch_size': 64, 'max_seq_len': 2048, 'n_eval': 500}, 'effective_config': {'base_model': 'Qwen/Qwen3.5-4B', 'arms': ['pe', 'trimuon'], 'lr_grid': [5e-05, 0.0001, 0.0002], 'runs': 6, 'steps': 80, 'batch_size': 64, 'max_seq_len': 2048, 'n_eval': 500}, 'primary_endpoint': 'mean per-token NLL on assistant tokens (no_robots test, 500)', 'source': 'probe_result_v1.json (warm timing pilot) and runs/*.jsonl', 'measurements': {'setup': {'seconds': 900, 'units': 1, 'planned_units': 1, 'samples': 0, 'gpu_count': None, 'allocation_source': None}, 'training': {'seconds': 1.19, 'units': 1, 'planned_units': 480, 'samples': 64, 'gpu_count': 4, 'allocation_source': 'fork config qwen35-4b-lora-32k trainer_gpus_per_node=4'}, 'evaluation': {'seconds': 3.84, 'units': 500, 'planned_units': 3000, 'samples': 500, 'gpu_count': 4, 'allocation_source': 'same trainer'}}, 'generation': None, 'recovery': {'seconds': 1200, 'gpu_count': None, 'allocation_source': None}, 'observed': {'six_run_sweep_wall_seconds': 600, 'trainer_gpu_hours_from_container_lifetimes': 2.2, 'preflight_gpu_hours': 0.15, 'note': "4xH100 trainer containers 17:34-17:46 (replaced after adapter-init fix) and 17:50-18:12; observed GPU use (~2.35 GPU-h incl. idle/cold start) exceeded the spec's 1.8 GPU-h budget; wall time ~56 min of 120"}}, 'recorded_at': 1793816220.634896}

feasibility_reported_verdict

refuted

feasibility_required

True

initialization_provenance

{'version': 1, 'summary': 'all 6 runs adapter_sha256=cb27f8c5c44d6a31e5c0df713ef62c37c76e37e3a7557bdd6986f22693d076a6 (server_response, before_first_update); restore_count=0; optimizer family verified every step via server codes', 'notes': "All six runs report the same pre-update adapter sha256. Initial-output logprob hashes on the fixed 4-item batch match for 4/6 runs (a8a72f25...); trimuon lr5e-5 and lr1e-4 differ although LoRA B=0 and the adapter hash matches, consistent with batch-composition-dependent forward numerics when the trainer packs concurrent clients' requests. No checkpoint restores are part of this spec, so restoration checks are not applicable."}

launch

{'experiment_id': 'exp_ff293ca443', 'context_status': 'legacy_unknown', 'superseded': False, 'mode': 'ordinary', 'purpose': 'Execute the selected saved experiment version.'}

progress_updated_at

1793816170.4581876

reported_verdict

refuted

restoration_validation

{'status': 'unverified', 'required': False, 'checks': [], 'reason': 'Initialization and optimizer evidence was not required by this saved spec.', 'applicability': 'unknown'}

started_at

1793812669.9046245

verdict_report

{'effect': 0.0005206562216542032, 'margin': 0.01, 'lower': 0.00025341916006382803, 'upper': 0.0007831968928146971, 'seeds': 1, 'confidence': 'weak', 'reasoning': 'Best-LR TriMuon eval NLL is 0.00052 nats above dimuon-lite (paired item-bootstrap 95% CI [0.00025, 0.00078]), so the required 0.01-nat improvement is not met; the step-time criterion is met (0.046 s vs 0.060 s median server optimizer step). Confidence is weak: one training seed, both arms selected the top LR of the grid (2e-4, grid edge), and the item bootstrap does not capture training-seed noise.', 'grid_edge': True}