diff --git a/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/ckpt.pt b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/ckpt.pt new file mode 100644 index 0000000000000000000000000000000000000000..fe308a82816efc78c0e8e684212e5ea97ccf7cbe --- /dev/null +++ b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/ckpt.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b06b2bc0f67daa03cb68684cfeecff25ebfedcea96b5a2f39afcb0acd942164 +size 23171663 diff --git a/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/config.json b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/config.json new file mode 100644 index 0000000000000000000000000000000000000000..051fb870bc1fe1a07b2ad2f9a91b92178a412f3b --- /dev/null +++ b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/config.json @@ -0,0 +1,36 @@ +{ + "size": "5m", + "name": "5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask", + "seed": 0, + "tokens": 400000000.0, + "batch_size": 64, + "block_size": 512, + "lr": 0.0006, + "eval_every": 4000, + "vocab": "bpe4096", + "pos": "none", + "reason": "step", + "reason_codec": "bytes", + "per_chunk": 3, + "docmask": true, + "compile": true, + "holdout": [ + "x.add.cmp", + "x.runmax.count", + "x.count.find", + "x.parity.rev", + "x.sumd.inc", + "x.inc.sumd" + ], + "dataset": "v11", + "vocab_size": 4096, + "n_layer": 6, + "n_head": 4, + "d_model": 256, + "loops": 1, + "params": 5787648, + "flops_per_token": 44163072, + "steps": 12207, + "n_docs": 1173825, + "stream_tokens": 215621900 +} \ No newline at end of file diff --git a/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/log.jsonl b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/log.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..756f5c26226477d996a66d7041e7c10ade117786 --- /dev/null +++ b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/log.jsonl @@ -0,0 +1,616 @@ +{"step": 0, "tokens": 0, "val_loss": 8.364864063262939, "val_bpb": 3.0153975970876363, "lr": 0.0, "elapsed": 0.47574305534362793, "peak_mem_gb": 1.1237632, "add8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5255405568.0}, "add9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5829525504.0}, "add10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6403645440.0}, "reach14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5299568640.0}, "reach16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5939403227.136}, "reach18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6493031497.728}, "parity24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8170168320.0}, "parity28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9406734336.0}, "parity32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10643300352.0}, "cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 2558278434.816}, "cmp9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 2381626146.816}, "cmp10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 2469952290.816}, "runmax12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4460470272.0}, "runmax14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5078753280.0}, "runmax16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5697036288.0}, "count16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5034590208.0}, "count18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5564547072.0}, "count20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6137607094.272}, "brackets16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5480460582.912}, "brackets18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6006177792.0}, "brackets20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6581357641.728}, "sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 3753861120.0}, "sumd9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4151328768.0}, "sumd10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4548796416.0}, "inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4107165696.0}, "inc9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4548796416.0}, "inc10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4990427136.0}, "add.cmp6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6211094446.08}, "add.cmp7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7018042097.664}, "add.cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7748675960.832}, "add.runmax6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7021928448.0}, "add.runmax7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7905189888.0}, "add.runmax8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8788451328.0}, "add.count6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6801113088.0}, "add.count7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7640211456.0}, "add.count8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8479309824.0}, "add.find6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7021928448.0}, "add.find7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7905189888.0}, "add.find8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8788451328.0}, "add.rev6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7331069952.0}, "add.rev7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8479309824.0}, "add.rev8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9671712768.0}, "add.sumd6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7330010038.272}, "add.sumd7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8303717449.728}, "add.sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9273185206.272}, "add.inc6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7640211456.0}, "add.inc7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8655962112.0}, "add.inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9671712768.0}, "runmax.cmp12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5432057856.0}, "runmax.cmp14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6050340864.0}, "runmax.cmp16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6668623872.0}, "runmax.runmax12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5520384000.0}, "runmax.runmax14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6138667008.0}, "runmax.runmax16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6756950016.0}, "runmax.count12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5564547072.0}, "runmax.count14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6182830080.0}, "runmax.count16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6801113088.0}, "runmax.find12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5520384000.0}, "runmax.find14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6138667008.0}, "runmax.find16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6756950016.0}, "runmax.rev12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5167079424.0}, "runmax.rev14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5785362432.0}, "runmax.rev16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6403645440.0}, "runmax.sumd12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5343731712.0}, "runmax.sumd14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5962014720.0}, "runmax.sumd16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6580297728.0}, "runmax.inc12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5387894784.0}, "runmax.inc14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6006177792.0}, "runmax.inc16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6624460800.0}, "count.cmp16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6006177792.0}, "count.cmp18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6536929591.296}, "count.cmp20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7066974781.44}, "count.runmax16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6094503936.0}, "count.runmax18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6624460800.0}, "count.runmax20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7348381876.224}, "count.count16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6138667008.0}, "count.count18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6668623872.0}, "count.count20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7726417772.544}, "count.find16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6094503936.0}, "count.find18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6624460800.0}, "count.find20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7154417664.0}, "count.rev16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5741199360.0}, "count.rev18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6271156224.0}, "count.rev20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6995077300.224}, "count.sumd16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5917851648.0}, "count.sumd18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6447808512.0}, "count.sumd20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7171729588.224}, "count.inc16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5940463140.864}, "count.inc18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6470420004.864}, "count.inc20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7258995818.496}, "parity.cmp16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6668623872.0}, "parity.cmp20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7905189888.0}, "parity.cmp24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9141755904.0}, "parity.runmax16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6756950016.0}, "parity.runmax20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7993516032.0}, "parity.runmax24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9230082048.0}, "parity.count16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6801113088.0}, "parity.count20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8037679104.0}, "parity.count24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9274245120.0}, "parity.find16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6756950016.0}, "parity.find20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7993516032.0}, "parity.find24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9230082048.0}, "parity.rev16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6403645440.0}, "parity.rev20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7640211456.0}, "parity.rev24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8876777472.0}, "parity.sumd16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6580297728.0}, "parity.sumd20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7816863744.0}, "parity.sumd24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9053429760.0}, "parity.inc16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6580297728.0}, "parity.inc20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7816863744.0}, "parity.inc24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9053429760.0}, "sumd.cmp6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4324624662.528}, "sumd.cmp7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4725448704.0}, "sumd.cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5120884850.688}, "sumd.runmax6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4327981056.0}, "sumd.runmax7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4725448704.0}, "sumd.runmax8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5122916352.0}, "sumd.count6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4327981056.0}, "sumd.count7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4725448704.0}, "sumd.count8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5122916352.0}, "sumd.find6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4327981056.0}, "sumd.find7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4725448704.0}, "sumd.find8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5122916352.0}, "sumd.rev6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 3974676480.0}, "sumd.rev7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4372144128.0}, "sumd.rev8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4769611776.0}, "sumd.sumd6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4239654912.0}, "sumd.sumd7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4637122560.0}, "sumd.sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5034590208.0}, "sumd.inc6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4283817984.0}, "sumd.inc7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 4681285632.0}, "sumd.inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5078753280.0}, "inc.cmp6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5523563741.184}, "inc.cmp7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6097683677.184}, "inc.cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6930422562.816}, "inc.runmax6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5829525504.0}, "inc.runmax7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6580297728.0}, "inc.runmax8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7331069952.0}, "inc.count6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5652873216.0}, "inc.count7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6359482368.0}, "inc.count8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7066091520.0}, "inc.find6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5829525504.0}, "inc.find7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6580297728.0}, "inc.find8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7331069952.0}, "inc.rev6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5917851648.0}, "inc.rev7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6889439232.0}, "inc.rev8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7905189888.0}, "inc.sumd6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6094503936.0}, "inc.sumd7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6933602304.0}, "inc.sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7772700672.0}, "inc.inc6": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6315319296.0}, "inc.inc7": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7198580736.0}, "inc.inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8081842176.0}, "x.add.cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9161894264.832}, "x.add.cmp9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9955769647.104}, "x.add.cmp10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10750704943.104}, "x.add.runmax8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10334158848.0}, "x.add.runmax9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11305746432.0}, "x.add.runmax10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12277334016.0}, "x.add.count8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9980854272.0}, "x.add.count9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10908278784.0}, "x.add.count10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11835703296.0}, "x.add.find8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10245832704.0}, "x.add.find9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11217420288.0}, "x.add.find10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12233170944.0}, "x.add.rev8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11084931072.0}, "x.add.rev9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12409823232.0}, "x.add.rev10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 13778878464.0}, "x.add.sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10730566582.272}, "x.add.sumd9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11835703296.0}, "x.add.sumd10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12895617024.0}, "x.add.inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11084931072.0}, "x.add.inc9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12189007872.0}, "x.add.inc10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 13293084672.0}, "x.runmax.cmp12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6889439232.0}, "x.runmax.cmp14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7596048384.0}, "x.runmax.cmp16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8302657536.0}, "x.runmax.runmax12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7110254592.0}, "x.runmax.runmax14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7816863744.0}, "x.runmax.runmax16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8523472896.0}, "x.runmax.count12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7110254592.0}, "x.runmax.count14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7816863744.0}, "x.runmax.count16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8523472896.0}, "x.runmax.find12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7021928448.0}, "x.runmax.find14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7728537600.0}, "x.runmax.find16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8435146752.0}, "x.runmax.rev12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6624460800.0}, "x.runmax.rev14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7331069952.0}, "x.runmax.rev16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8037679104.0}, "x.runmax.sumd12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6845276160.0}, "x.runmax.sumd14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7551885312.0}, "x.runmax.sumd16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8258494464.0}, "x.runmax.inc12": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6845276160.0}, "x.runmax.inc14": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7551885312.0}, "x.runmax.inc16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8258494464.0}, "x.count.cmp16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7640211456.0}, "x.count.cmp18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8259289399.296}, "x.count.cmp20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8877660733.44}, "x.count.runmax16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7861026816.0}, "x.count.runmax18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8479309824.0}, "x.count.runmax20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9291557044.224}, "x.count.count16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7861026816.0}, "x.count.count18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8479309824.0}, "x.count.count20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9625429868.544}, "x.count.find16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7772700672.0}, "x.count.find18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8390983680.0}, "x.count.find20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9009266688.0}, "x.count.rev16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7375233024.0}, "x.count.rev18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7993516032.0}, "x.count.rev20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8805763252.224}, "x.count.sumd16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7596048384.0}, "x.count.sumd18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8214331392.0}, "x.count.sumd20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9026578612.224}, "x.count.inc16": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7574496804.864}, "x.count.inc18": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8192779812.864}, "x.count.inc20": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9069681770.496}, "x.parity.cmp24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11129094144.0}, "x.parity.cmp28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12542312448.0}, "x.parity.cmp32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 13955530752.0}, "x.parity.runmax24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11349909504.0}, "x.parity.runmax28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12763127808.0}, "x.parity.runmax32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 14176346112.0}, "x.parity.count24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11349909504.0}, "x.parity.count28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12763127808.0}, "x.parity.count32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 14176346112.0}, "x.parity.find24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11261583360.0}, "x.parity.find28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12674801664.0}, "x.parity.find32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 14088019968.0}, "x.parity.rev24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10864115712.0}, "x.parity.rev28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12277334016.0}, "x.parity.rev32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 13690552320.0}, "x.parity.sumd24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11084931072.0}, "x.parity.sumd28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12498149376.0}, "x.parity.sumd32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 13911367680.0}, "x.parity.inc24": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11040768000.0}, "x.parity.inc28": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 12453986304.0}, "x.parity.inc32": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 13867204608.0}, "x.sumd.cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6224961650.688}, "x.sumd.cmp9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6668623872.0}, "x.sumd.cmp10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7110254592.0}, "x.sumd.runmax8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6359482368.0}, "x.sumd.runmax9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6801113088.0}, "x.sumd.runmax10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7242743808.0}, "x.sumd.count8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6315319296.0}, "x.sumd.count9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6756950016.0}, "x.sumd.count10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7198580736.0}, "x.sumd.find8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6271156224.0}, "x.sumd.find9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6712786944.0}, "x.sumd.find10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7154417664.0}, "x.sumd.rev8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 5873688576.0}, "x.sumd.rev9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6315319296.0}, "x.sumd.rev10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6756950016.0}, "x.sumd.sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6182830080.0}, "x.sumd.sumd9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6624460800.0}, "x.sumd.sumd10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7066091520.0}, "x.sumd.inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6182830080.0}, "x.sumd.inc9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 6624460800.0}, "x.sumd.inc10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7066091520.0}, "x.inc.cmp8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 7946173218.816}, "x.inc.cmp9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8299477794.816}, "x.inc.cmp10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8917760802.816}, "x.inc.runmax8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8479309824.0}, "x.inc.runmax9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9274245120.0}, "x.inc.runmax10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10069180416.0}, "x.inc.count8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8170168320.0}, "x.inc.count9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8920940544.0}, "x.inc.count10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9671712768.0}, "x.inc.find8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8390983680.0}, "x.inc.find9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9185918976.0}, "x.inc.find10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9980854272.0}, "x.inc.rev8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8920940544.0}, "x.inc.rev9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10025017344.0}, "x.inc.rev10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 11173257216.0}, "x.inc.sumd8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 8832614400.0}, "x.inc.sumd9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9715875840.0}, "x.inc.sumd10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10599137280.0}, "x.inc.inc8": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 9097592832.0}, "x.inc.inc9": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10025017344.0}, "x.inc.inc10": {"acc": 0.0, "reasoning_bytes": 0.0, "flops_per_problem": 10952441856.0}} +{"step": 0, "train_loss": 8.341028213500977} +{"step": 20, "train_loss": 7.19752311706543} +{"step": 40, "train_loss": 6.521101474761963} +{"step": 60, "train_loss": 5.785541534423828} +{"step": 80, "train_loss": 4.772795677185059} +{"step": 100, "train_loss": 4.087849140167236} +{"step": 120, "train_loss": 3.6280946731567383} +{"step": 140, "train_loss": 2.9904656410217285} +{"step": 160, "train_loss": 2.946800708770752} +{"step": 180, "train_loss": 2.5865368843078613} +{"step": 200, "train_loss": 2.3900692462921143} +{"step": 220, "train_loss": 2.447683095932007} +{"step": 240, "train_loss": 2.387730598449707} +{"step": 260, "train_loss": 2.2213547229766846} +{"step": 280, "train_loss": 2.171464443206787} +{"step": 300, "train_loss": 2.1918883323669434} +{"step": 320, "train_loss": 2.1690711975097656} +{"step": 340, "train_loss": 2.1443350315093994} +{"step": 360, "train_loss": 2.0338988304138184} +{"step": 380, "train_loss": 2.198914051055908} +{"step": 400, "train_loss": 2.1205520629882812} +{"step": 420, "train_loss": 2.0761301517486572} +{"step": 440, "train_loss": 1.970961570739746} +{"step": 460, "train_loss": 1.8446305990219116} +{"step": 480, "train_loss": 1.8906302452087402} +{"step": 500, "train_loss": 2.035715103149414} +{"step": 520, "train_loss": 1.8665324449539185} +{"step": 540, "train_loss": 1.9358965158462524} +{"step": 560, "train_loss": 1.8654439449310303} +{"step": 580, "train_loss": 1.9571294784545898} +{"step": 600, "train_loss": 1.7516014575958252} +{"step": 620, "train_loss": 1.8118093013763428} +{"step": 640, "train_loss": 1.828635334968567} +{"step": 660, "train_loss": 1.6980539560317993} +{"step": 680, "train_loss": 1.6501997709274292} +{"step": 700, "train_loss": 1.790748119354248} +{"step": 720, "train_loss": 1.6633739471435547} +{"step": 740, "train_loss": 1.7010042667388916} +{"step": 760, "train_loss": 1.6405447721481323} +{"step": 780, "train_loss": 1.6363482475280762} +{"step": 800, "train_loss": 1.6865503787994385} +{"step": 820, "train_loss": 1.6912732124328613} +{"step": 840, "train_loss": 1.5796594619750977} +{"step": 860, "train_loss": 1.5111474990844727} +{"step": 880, "train_loss": 1.575636386871338} +{"step": 900, "train_loss": 1.544931411743164} +{"step": 920, "train_loss": 1.577256441116333} +{"step": 940, "train_loss": 1.6572535037994385} +{"step": 960, "train_loss": 1.5153539180755615} +{"step": 980, "train_loss": 1.574941873550415} +{"step": 1000, "train_loss": 1.4166131019592285} +{"step": 1020, "train_loss": 1.5123904943466187} +{"step": 1040, "train_loss": 1.5046405792236328} +{"step": 1060, "train_loss": 1.4856114387512207} +{"step": 1080, "train_loss": 1.4078726768493652} +{"step": 1100, "train_loss": 1.4127089977264404} +{"step": 1120, "train_loss": 1.4954478740692139} +{"step": 1140, "train_loss": 1.3268373012542725} +{"step": 1160, "train_loss": 1.5535786151885986} +{"step": 1180, "train_loss": 1.415763020515442} +{"step": 1200, "train_loss": 1.5271496772766113} +{"step": 1220, "train_loss": 1.4487859010696411} +{"step": 1240, "train_loss": 1.4423792362213135} +{"step": 1260, "train_loss": 1.3348459005355835} +{"step": 1280, "train_loss": 1.4007320404052734} +{"step": 1300, "train_loss": 1.4651050567626953} +{"step": 1320, "train_loss": 1.394850730895996} +{"step": 1340, "train_loss": 1.4090827703475952} +{"step": 1360, "train_loss": 1.400513768196106} +{"step": 1380, "train_loss": 1.3893346786499023} +{"step": 1400, "train_loss": 1.400663137435913} +{"step": 1420, "train_loss": 1.3289613723754883} +{"step": 1440, "train_loss": 1.4721646308898926} +{"step": 1460, "train_loss": 1.437692642211914} +{"step": 1480, "train_loss": 1.324838638305664} +{"step": 1500, "train_loss": 1.3155584335327148} +{"step": 1520, "train_loss": 1.3097808361053467} +{"step": 1540, "train_loss": 1.3713781833648682} +{"step": 1560, "train_loss": 1.213754415512085} +{"step": 1580, "train_loss": 1.3109508752822876} +{"step": 1600, "train_loss": 1.296341061592102} +{"step": 1620, "train_loss": 1.370373010635376} +{"step": 1640, "train_loss": 1.3509511947631836} +{"step": 1660, "train_loss": 1.390860915184021} +{"step": 1680, "train_loss": 1.315922737121582} +{"step": 1700, "train_loss": 1.4304370880126953} +{"step": 1720, "train_loss": 1.2848410606384277} +{"step": 1740, "train_loss": 1.3144042491912842} +{"step": 1760, "train_loss": 1.319590449333191} +{"step": 1780, "train_loss": 1.33492112159729} +{"step": 1800, "train_loss": 1.2455685138702393} +{"step": 1820, "train_loss": 1.27618408203125} +{"step": 1840, "train_loss": 1.2916325330734253} +{"step": 1860, "train_loss": 1.417574405670166} +{"step": 1880, "train_loss": 1.2179768085479736} +{"step": 1900, "train_loss": 1.193871021270752} +{"step": 1920, "train_loss": 1.283959984779358} +{"step": 1940, "train_loss": 1.2946287393569946} +{"step": 1960, "train_loss": 1.3288716077804565} +{"step": 1980, "train_loss": 1.2688201665878296} +{"step": 2000, "train_loss": 1.3369488716125488} +{"step": 2020, "train_loss": 1.2689834833145142} +{"step": 2040, "train_loss": 1.2439866065979004} +{"step": 2060, "train_loss": 1.3320159912109375} +{"step": 2080, "train_loss": 1.259772777557373} +{"step": 2100, "train_loss": 1.2045652866363525} +{"step": 2120, "train_loss": 1.3108878135681152} +{"step": 2140, "train_loss": 1.2572170495986938} +{"step": 2160, "train_loss": 1.2045025825500488} +{"step": 2180, "train_loss": 1.2644708156585693} +{"step": 2200, "train_loss": 1.2545497417449951} +{"step": 2220, "train_loss": 1.2770500183105469} +{"step": 2240, "train_loss": 1.249247431755066} +{"step": 2260, "train_loss": 1.2651602029800415} +{"step": 2280, "train_loss": 1.2727469205856323} +{"step": 2300, "train_loss": 1.2473597526550293} +{"step": 2320, "train_loss": 1.2113217115402222} +{"step": 2340, "train_loss": 1.2158710956573486} +{"step": 2360, "train_loss": 1.3278932571411133} +{"step": 2380, "train_loss": 1.209028959274292} +{"step": 2400, "train_loss": 1.3148486614227295} +{"step": 2420, "train_loss": 1.1820309162139893} +{"step": 2440, "train_loss": 1.2140953540802002} +{"step": 2460, "train_loss": 1.3275182247161865} +{"step": 2480, "train_loss": 1.2562612295150757} +{"step": 2500, "train_loss": 1.147978663444519} +{"step": 2520, "train_loss": 1.1242847442626953} +{"step": 2540, "train_loss": 1.2221171855926514} +{"step": 2560, "train_loss": 1.1754194498062134} +{"step": 2580, "train_loss": 1.2085431814193726} +{"step": 2600, "train_loss": 1.2409758567810059} +{"step": 2620, "train_loss": 1.2047388553619385} +{"step": 2640, "train_loss": 1.2392022609710693} +{"step": 2660, "train_loss": 1.2067480087280273} +{"step": 2680, "train_loss": 1.2139949798583984} +{"step": 2700, "train_loss": 1.230457067489624} +{"step": 2720, "train_loss": 1.2348672151565552} +{"step": 2740, "train_loss": 1.305181622505188} +{"step": 2760, "train_loss": 1.2712522745132446} +{"step": 2780, "train_loss": 1.2740364074707031} +{"step": 2800, "train_loss": 1.1791739463806152} +{"step": 2820, "train_loss": 1.2191057205200195} +{"step": 2840, "train_loss": 1.216299295425415} +{"step": 2860, "train_loss": 1.2585854530334473} +{"step": 2880, "train_loss": 1.193520188331604} +{"step": 2900, "train_loss": 1.197994351387024} +{"step": 2920, "train_loss": 1.1950414180755615} +{"step": 2940, "train_loss": 1.278044581413269} +{"step": 2960, "train_loss": 1.2191532850265503} +{"step": 2980, "train_loss": 1.182539701461792} +{"step": 3000, "train_loss": 1.1956818103790283} +{"step": 3020, "train_loss": 1.1893174648284912} +{"step": 3040, "train_loss": 1.2124114036560059} +{"step": 3060, "train_loss": 1.2460321187973022} +{"step": 3080, "train_loss": 1.2563178539276123} +{"step": 3100, "train_loss": 1.1851308345794678} +{"step": 3120, "train_loss": 1.2066636085510254} +{"step": 3140, "train_loss": 1.1123487949371338} +{"step": 3160, "train_loss": 1.153381109237671} +{"step": 3180, "train_loss": 1.1576229333877563} +{"step": 3200, "train_loss": 1.1861515045166016} +{"step": 3220, "train_loss": 1.3114233016967773} +{"step": 3240, "train_loss": 1.245585322380066} +{"step": 3260, "train_loss": 1.1850320100784302} +{"step": 3280, "train_loss": 1.0250940322875977} +{"step": 3300, "train_loss": 1.1709132194519043} +{"step": 3320, "train_loss": 1.1954030990600586} +{"step": 3340, "train_loss": 1.2151044607162476} +{"step": 3360, "train_loss": 1.1605983972549438} +{"step": 3380, "train_loss": 1.2455145120620728} +{"step": 3400, "train_loss": 1.2160767316818237} +{"step": 3420, "train_loss": 1.2203665971755981} +{"step": 3440, "train_loss": 1.130004644393921} +{"step": 3460, "train_loss": 1.1747896671295166} +{"step": 3480, "train_loss": 1.2046053409576416} +{"step": 3500, "train_loss": 1.1294867992401123} +{"step": 3520, "train_loss": 1.1878527402877808} +{"step": 3540, "train_loss": 1.2225158214569092} +{"step": 3560, "train_loss": 1.2644050121307373} +{"step": 3580, "train_loss": 1.1069715023040771} +{"step": 3600, "train_loss": 1.1679145097732544} +{"step": 3620, "train_loss": 1.1826354265213013} +{"step": 3640, "train_loss": 1.193589448928833} +{"step": 3660, "train_loss": 1.1409196853637695} +{"step": 3680, "train_loss": 1.1214988231658936} +{"step": 3700, "train_loss": 1.14332914352417} +{"step": 3720, "train_loss": 1.0917906761169434} +{"step": 3740, "train_loss": 1.1685030460357666} +{"step": 3760, "train_loss": 1.1614265441894531} +{"step": 3780, "train_loss": 1.1860804557800293} +{"step": 3800, "train_loss": 1.1517152786254883} +{"step": 3820, "train_loss": 1.0044857263565063} +{"step": 3840, "train_loss": 1.1561603546142578} +{"step": 3860, "train_loss": 1.1707843542099} +{"step": 3880, "train_loss": 1.1636545658111572} +{"step": 3900, "train_loss": 1.1480566263198853} +{"step": 3920, "train_loss": 1.138697624206543} +{"step": 3940, "train_loss": 1.0820367336273193} +{"step": 3960, "train_loss": 1.112986445426941} +{"step": 3980, "train_loss": 1.142602801322937} +{"step": 4000, "tokens": 131072000, "val_loss": 2.0936762928962707, "val_bpb": 0.7547362891891521, "lr": 0.0004789669757536829, "elapsed": 293.159592628479, "peak_mem_gb": 2.254330368, "add8": {"acc": 0.024, "reasoning_bytes": 69.46, "flops_per_problem": 4462590099.456}, "add9": {"acc": 0.0, "reasoning_bytes": 79.008, "flops_per_problem": 5015070130.176}, "add10": {"acc": 0.0, "reasoning_bytes": 88.388, "flops_per_problem": 5561897287.68}, "reach14": {"acc": 0.488, "reasoning_bytes": 7.014, "flops_per_problem": 4594196054.016}, "reach16": {"acc": 0.496, "reasoning_bytes": 7.658, "flops_per_problem": 5152417284.096}, "reach18": {"acc": 0.488, "reasoning_bytes": 8.356, "flops_per_problem": 5712758341.632}, "parity24": {"acc": 0.0, "reasoning_bytes": 147.0, "flops_per_problem": 8170168320.0}, "parity28": {"acc": 0.0, "reasoning_bytes": 171.024, "flops_per_problem": 9406734336.0}, "parity32": {"acc": 0.0, "reasoning_bytes": 196.188, "flops_per_problem": 10643300352.0}, "cmp8": {"acc": 0.978, "reasoning_bytes": 5.908, "flops_per_problem": 1365257207.808}, "cmp9": {"acc": 0.974, "reasoning_bytes": 5.872, "flops_per_problem": 1451993481.216}, "cmp10": {"acc": 0.916, "reasoning_bytes": 7.64, "flops_per_problem": 1618399936.512}, "runmax12": {"acc": 0.972, "reasoning_bytes": 76.98, "flops_per_problem": 4416748830.72}, "runmax14": {"acc": 0.932, "reasoning_bytes": 88.984, "flops_per_problem": 5037239992.32}, "runmax16": {"acc": 0.75, "reasoning_bytes": 100.92, "flops_per_problem": 5660734242.816}, "count16": {"acc": 0.3, "reasoning_bytes": 85.028, "flops_per_problem": 4965430837.248}, "count18": {"acc": 0.25, "reasoning_bytes": 95.17, "flops_per_problem": 5501393879.04}, "count20": {"acc": 0.162, "reasoning_bytes": 105.652, "flops_per_problem": 6073129009.152}, "brackets16": {"acc": 0.838, "reasoning_bytes": 44.85, "flops_per_problem": 3261354541.056}, "brackets18": {"acc": 0.864, "reasoning_bytes": 47.2, "flops_per_problem": 3455760384.0}, "brackets20": {"acc": 0.794, "reasoning_bytes": 54.31, "flops_per_problem": 3857997643.776}, "sumd8": {"acc": 0.664, "reasoning_bytes": 59.19, "flops_per_problem": 3406209417.216}, "sumd9": {"acc": 0.312, "reasoning_bytes": 66.668, "flops_per_problem": 3782037159.936}, "sumd10": {"acc": 0.06, "reasoning_bytes": 73.854, "flops_per_problem": 4146647482.368}, "inc8": {"acc": 0.226, "reasoning_bytes": 57.048, "flops_per_problem": 3502308261.888}, "inc9": {"acc": 0.0, "reasoning_bytes": 63.0, "flops_per_problem": 3842187264.0}, "inc10": {"acc": 0.0, "reasoning_bytes": 65.896, "flops_per_problem": 4030586929.152}, "add.cmp6": {"acc": 0.916, "reasoning_bytes": 90.62, "flops_per_problem": 5443628580.864}, "add.cmp7": {"acc": 0.788, "reasoning_bytes": 97.686, "flops_per_problem": 5886230888.448}, "add.cmp8": {"acc": 0.8, "reasoning_bytes": 106.48, "flops_per_problem": 6408061747.2}, "add.runmax6": {"acc": 0.94, "reasoning_bytes": 119.666, "flops_per_problem": 6521384189.952}, "add.runmax7": {"acc": 0.776, "reasoning_bytes": 124.558, "flops_per_problem": 6825667756.032}, "add.runmax8": {"acc": 0.64, "reasoning_bytes": 139.2, "flops_per_problem": 7560717926.4}, "add.count6": {"acc": 0.638, "reasoning_bytes": 107.65, "flops_per_problem": 6034883788.8}, "add.count7": {"acc": 0.404, "reasoning_bytes": 115.886, "flops_per_problem": 6486936993.792}, "add.count8": {"acc": 0.44, "reasoning_bytes": 134.916, "flops_per_problem": 7415686397.952}, "add.find6": {"acc": 0.666, "reasoning_bytes": 102.55, "flops_per_problem": 5765489049.6}, "add.find7": {"acc": 0.372, "reasoning_bytes": 112.056, "flops_per_problem": 6273629356.032}, "add.find8": {"acc": 0.55, "reasoning_bytes": 124.948, "flops_per_problem": 6934838870.016}, "add.rev6": {"acc": 0.49, "reasoning_bytes": 124.616, "flops_per_problem": 6794135322.624}, "add.rev7": {"acc": 0.094, "reasoning_bytes": 130.494, "flops_per_problem": 7209003220.992}, "add.rev8": {"acc": 0.0, "reasoning_bytes": 150.61, "flops_per_problem": 8243037388.8}, "add.sumd6": {"acc": 0.256, "reasoning_bytes": 129.37, "flops_per_problem": 6901098283.008}, "add.sumd7": {"acc": 0.07, "reasoning_bytes": 131.972, "flops_per_problem": 7108399742.976}, "add.sumd8": {"acc": 0.046, "reasoning_bytes": 149.11, "flops_per_problem": 7954210897.92}, "add.inc6": {"acc": 0.694, "reasoning_bytes": 127.494, "flops_per_problem": 6973349068.8}, "add.inc7": {"acc": 0.066, "reasoning_bytes": 133.298, "flops_per_problem": 7330893299.712}, "add.inc8": {"acc": 0.002, "reasoning_bytes": 150.688, "flops_per_problem": 8227227009.024}, "runmax.cmp12": {"acc": 0.99, "reasoning_bytes": 89.564, "flops_per_problem": 5236150468.608}, "runmax.cmp14": {"acc": 0.964, "reasoning_bytes": 99.288, "flops_per_problem": 5757186392.064}, "runmax.cmp16": {"acc": 0.97, "reasoning_bytes": 111.39, "flops_per_problem": 6380327337.984}, "runmax.runmax12": {"acc": 0.992, "reasoning_bytes": 90.472, "flops_per_problem": 5320413609.984}, "runmax.runmax14": {"acc": 0.98, "reasoning_bytes": 102.04, "flops_per_problem": 5919618170.88}, "runmax.runmax16": {"acc": 0.964, "reasoning_bytes": 112.4, "flops_per_problem": 6465385414.656}, "runmax.count12": {"acc": 1.0, "reasoning_bytes": 91.036, "flops_per_problem": 5389484654.592}, "runmax.count14": {"acc": 1.0, "reasoning_bytes": 101.224, "flops_per_problem": 5927744176.128}, "runmax.count16": {"acc": 1.0, "reasoning_bytes": 112.24, "flops_per_problem": 6502570721.28}, "runmax.find12": {"acc": 0.998, "reasoning_bytes": 93.736, "flops_per_problem": 5464561876.992}, "runmax.find14": {"acc": 0.992, "reasoning_bytes": 104.032, "flops_per_problem": 6007591010.304}, "runmax.find16": {"acc": 0.976, "reasoning_bytes": 112.764, "flops_per_problem": 6481460772.864}, "runmax.rev12": {"acc": 0.996, "reasoning_bytes": 87.416, "flops_per_problem": 5052962045.952}, "runmax.rev14": {"acc": 0.982, "reasoning_bytes": 97.58, "flops_per_problem": 5590161653.76}, "runmax.rev16": {"acc": 0.956, "reasoning_bytes": 107.7, "flops_per_problem": 6125241434.112}, "runmax.sumd12": {"acc": 0.996, "reasoning_bytes": 91.676, "flops_per_problem": 5285259804.672}, "runmax.sumd14": {"acc": 0.986, "reasoning_bytes": 102.152, "flops_per_problem": 5836238290.944}, "runmax.sumd16": {"acc": 0.972, "reasoning_bytes": 111.106, "flops_per_problem": 6320000581.632}, "runmax.inc12": {"acc": 0.996, "reasoning_bytes": 91.968, "flops_per_problem": 5285171478.528}, "runmax.inc14": {"acc": 0.972, "reasoning_bytes": 103.576, "flops_per_problem": 5890470543.36}, "runmax.inc16": {"acc": 0.876, "reasoning_bytes": 113.846, "flops_per_problem": 6450458296.32}, "count.cmp16": {"acc": 0.558, "reasoning_bytes": 98.836, "flops_per_problem": 5887909085.184}, "count.cmp18": {"acc": 0.624, "reasoning_bytes": 108.558, "flops_per_problem": 6400730677.248}, "count.cmp20": {"acc": 0.624, "reasoning_bytes": 118.094, "flops_per_problem": 6903836393.472}, "count.runmax16": {"acc": 0.426, "reasoning_bytes": 99.946, "flops_per_problem": 5996373590.016}, "count.runmax18": {"acc": 0.396, "reasoning_bytes": 109.91, "flops_per_problem": 6519352688.64}, "count.runmax20": {"acc": 0.428, "reasoning_bytes": 120.934, "flops_per_problem": 7070066196.48}, "count.count16": {"acc": 0.564, "reasoning_bytes": 99.698, "flops_per_problem": 6026669457.408}, "count.count18": {"acc": 0.552, "reasoning_bytes": 109.996, "flops_per_problem": 6557862887.424}, "count.count20": {"acc": 0.976, "reasoning_bytes": 121.408, "flops_per_problem": 7128979734.528}, "count.find16": {"acc": 0.514, "reasoning_bytes": 99.842, "flops_per_problem": 5985244495.872}, "count.find18": {"acc": 0.498, "reasoning_bytes": 109.986, "flops_per_problem": 6519529340.928}, "count.find20": {"acc": 0.558, "reasoning_bytes": 119.726, "flops_per_problem": 7025549819.904}, "count.rev16": {"acc": 0.462, "reasoning_bytes": 95.436, "flops_per_problem": 5635561291.776}, "count.rev18": {"acc": 0.384, "reasoning_bytes": 105.326, "flops_per_problem": 6162250088.448}, "count.rev20": {"acc": 0.422, "reasoning_bytes": 115.954, "flops_per_problem": 6715966685.184}, "count.sumd16": {"acc": 0.38, "reasoning_bytes": 98.354, "flops_per_problem": 5827758981.12}, "count.sumd18": {"acc": 0.364, "reasoning_bytes": 108.048, "flops_per_problem": 6341817139.2}, "count.sumd20": {"acc": 0.406, "reasoning_bytes": 118.992, "flops_per_problem": 6903748067.328}, "count.inc16": {"acc": 0.378, "reasoning_bytes": 98.47, "flops_per_problem": 5840124641.28}, "count.inc18": {"acc": 0.328, "reasoning_bytes": 108.26, "flops_per_problem": 6370699788.288}, "count.inc20": {"acc": 0.37, "reasoning_bytes": 120.368, "flops_per_problem": 6958598602.752}, "parity.cmp16": {"acc": 0.934, "reasoning_bytes": 116.656, "flops_per_problem": 6624814104.576}, "parity.cmp20": {"acc": 0.258, "reasoning_bytes": 134.208, "flops_per_problem": 7893089206.272}, "parity.cmp24": {"acc": 0.026, "reasoning_bytes": 159.848, "flops_per_problem": 9140607664.128}, "parity.runmax16": {"acc": 0.652, "reasoning_bytes": 116.536, "flops_per_problem": 6718086512.64}, "parity.runmax20": {"acc": 0.048, "reasoning_bytes": 132.348, "flops_per_problem": 7989453029.376}, "parity.runmax24": {"acc": 0.076, "reasoning_bytes": 160.244, "flops_per_problem": 9229728743.424}, "parity.count16": {"acc": 0.984, "reasoning_bytes": 117.56, "flops_per_problem": 6745997574.144}, "parity.count20": {"acc": 0.568, "reasoning_bytes": 137.284, "flops_per_problem": 8012594479.104}, "parity.count24": {"acc": 0.052, "reasoning_bytes": 157.554, "flops_per_problem": 9272655249.408}, "parity.find16": {"acc": 0.93, "reasoning_bytes": 117.78, "flops_per_problem": 6713670205.44}, "parity.find20": {"acc": 0.34, "reasoning_bytes": 135.334, "flops_per_problem": 7976115781.632}, "parity.find24": {"acc": 0.03, "reasoning_bytes": 157.886, "flops_per_problem": 9228933808.128}, "parity.rev16": {"acc": 0.588, "reasoning_bytes": 112.64, "flops_per_problem": 6362132152.32}, "parity.rev20": {"acc": 0.064, "reasoning_bytes": 132.056, "flops_per_problem": 7633940299.776}, "parity.rev24": {"acc": 0.004, "reasoning_bytes": 159.116, "flops_per_problem": 8876424167.424}, "parity.sumd16": {"acc": 0.474, "reasoning_bytes": 113.18, "flops_per_problem": 6550090186.752}, "parity.sumd20": {"acc": 0.012, "reasoning_bytes": 132.75, "flops_per_problem": 7815892156.416}, "parity.sumd24": {"acc": 0.21, "reasoning_bytes": 162.16, "flops_per_problem": 9053429760.0}, "parity.inc16": {"acc": 0.412, "reasoning_bytes": 117.916, "flops_per_problem": 6557156278.272}, "parity.inc20": {"acc": 0.0, "reasoning_bytes": 139.332, "flops_per_problem": 7816687091.712}, "parity.inc24": {"acc": 0.0, "reasoning_bytes": 165.316, "flops_per_problem": 9053429760.0}, "sumd.cmp6": {"acc": 0.98, "reasoning_bytes": 62.052, "flops_per_problem": 3711817875.456}, "sumd.cmp7": {"acc": 0.958, "reasoning_bytes": 70.068, "flops_per_problem": 4110168784.896}, "sumd.cmp8": {"acc": 0.94, "reasoning_bytes": 75.982, "flops_per_problem": 4415423938.56}, "sumd.runmax6": {"acc": 0.992, "reasoning_bytes": 67.542, "flops_per_problem": 3954449793.024}, "sumd.runmax7": {"acc": 0.988, "reasoning_bytes": 75.476, "flops_per_problem": 4349002678.272}, "sumd.runmax8": {"acc": 0.906, "reasoning_bytes": 83.314, "flops_per_problem": 4739315908.608}, "sumd.count6": {"acc": 0.996, "reasoning_bytes": 66.428, "flops_per_problem": 3949415202.816}, "sumd.count7": {"acc": 0.956, "reasoning_bytes": 74.444, "flops_per_problem": 4347589459.968}, "sumd.count8": {"acc": 0.854, "reasoning_bytes": 82.108, "flops_per_problem": 4730306641.92}, "sumd.find6": {"acc": 0.992, "reasoning_bytes": 66.23, "flops_per_problem": 3896507842.56}, "sumd.find7": {"acc": 0.982, "reasoning_bytes": 74.468, "flops_per_problem": 4304486301.696}, "sumd.find8": {"acc": 0.908, "reasoning_bytes": 82.194, "flops_per_problem": 4689853267.968}, "sumd.rev6": {"acc": 0.988, "reasoning_bytes": 61.544, "flops_per_problem": 3601145217.024}, "sumd.rev7": {"acc": 0.968, "reasoning_bytes": 69.48, "flops_per_problem": 3995874754.56}, "sumd.rev8": {"acc": 0.858, "reasoning_bytes": 77.496, "flops_per_problem": 4393872359.424}, "sumd.sumd6": {"acc": 0.974, "reasoning_bytes": 65.744, "flops_per_problem": 3795551059.968}, "sumd.sumd7": {"acc": 0.89, "reasoning_bytes": 73.732, "flops_per_problem": 4195138535.424}, "sumd.sumd8": {"acc": 0.736, "reasoning_bytes": 81.594, "flops_per_problem": 4586688331.776}, "sumd.inc6": {"acc": 0.964, "reasoning_bytes": 68.568, "flops_per_problem": 3913819766.784}, "sumd.inc7": {"acc": 0.91, "reasoning_bytes": 76.63, "flops_per_problem": 4312523980.8}, "sumd.inc8": {"acc": 0.578, "reasoning_bytes": 85.456, "flops_per_problem": 4748060196.864}, "inc.cmp6": {"acc": 0.978, "reasoning_bytes": 77.89, "flops_per_problem": 4543938478.08}, "inc.cmp7": {"acc": 0.924, "reasoning_bytes": 85.39, "flops_per_problem": 4963487662.08}, "inc.cmp8": {"acc": 0.914, "reasoning_bytes": 90.59, "flops_per_problem": 5281461780.48}, "inc.runmax6": {"acc": 1.0, "reasoning_bytes": 102.66, "flops_per_problem": 5461205483.52}, "inc.runmax7": {"acc": 0.976, "reasoning_bytes": 111.684, "flops_per_problem": 5903896117.248}, "inc.runmax8": {"acc": 0.912, "reasoning_bytes": 120.138, "flops_per_problem": 6321413799.936}, "inc.count6": {"acc": 0.86, "reasoning_bytes": 97.766, "flops_per_problem": 5289234481.152}, "inc.count7": {"acc": 0.706, "reasoning_bytes": 101.326, "flops_per_problem": 5490618089.472}, "inc.count8": {"acc": 0.67, "reasoning_bytes": 114.124, "flops_per_problem": 6099980156.928}, "inc.find6": {"acc": 0.86, "reasoning_bytes": 90.8, "flops_per_problem": 4937431449.6}, "inc.find7": {"acc": 0.752, "reasoning_bytes": 93.758, "flops_per_problem": 5112228888.576}, "inc.find8": {"acc": 0.716, "reasoning_bytes": 105.09, "flops_per_problem": 5656847892.48}, "inc.rev6": {"acc": 0.75, "reasoning_bytes": 100.96, "flops_per_problem": 5462442049.536}, "inc.rev7": {"acc": 0.294, "reasoning_bytes": 111.13, "flops_per_problem": 5980474884.096}, "inc.rev8": {"acc": 0.012, "reasoning_bytes": 125.158, "flops_per_problem": 6680194596.864}, "inc.sumd6": {"acc": 0.482, "reasoning_bytes": 108.274, "flops_per_problem": 5658261110.784}, "inc.sumd7": {"acc": 0.124, "reasoning_bytes": 114.626, "flops_per_problem": 5986657714.176}, "inc.sumd8": {"acc": 0.136, "reasoning_bytes": 129.56, "flops_per_problem": 6693178540.032}, "inc.inc6": {"acc": 0.442, "reasoning_bytes": 103.58, "flops_per_problem": 5569405009.92}, "inc.inc7": {"acc": 0.048, "reasoning_bytes": 113.95, "flops_per_problem": 6097948655.616}, "inc.inc8": {"acc": 0.004, "reasoning_bytes": 131.022, "flops_per_problem": 6939431829.504}, "x.add.cmp8": {"acc": 0.646, "reasoning_bytes": 149.02, "flops_per_problem": 8375614930.944}, "x.add.cmp9": {"acc": 0.62, "reasoning_bytes": 164.28, "flops_per_problem": 9201552703.488}, "x.add.cmp10": {"acc": 0.53, "reasoning_bytes": 176.116, "flops_per_problem": 9922382364.672}, "x.add.runmax8": {"acc": 0.754, "reasoning_bytes": 191.47, "flops_per_problem": 9956387930.112}, "x.add.runmax9": {"acc": 0.582, "reasoning_bytes": 205.522, "flops_per_problem": 10665381888.0}, "x.add.runmax10": {"acc": 0.602, "reasoning_bytes": 217.262, "flops_per_problem": 11272889106.432}, "x.add.count8": {"acc": 0.516, "reasoning_bytes": 183.22, "flops_per_problem": 9637265571.84}, "x.add.count9": {"acc": 0.44, "reasoning_bytes": 191.66, "flops_per_problem": 10098151391.232}, "x.add.count10": {"acc": 0.346, "reasoning_bytes": 203.624, "flops_per_problem": 10715021180.928}, "x.add.find8": {"acc": 0.498, "reasoning_bytes": 166.662, "flops_per_problem": 8861850353.664}, "x.add.find9": {"acc": 0.306, "reasoning_bytes": 179.05, "flops_per_problem": 9501508288.512}, "x.add.find10": {"acc": 0.238, "reasoning_bytes": 194.332, "flops_per_problem": 10281074835.456}, "x.add.rev8": {"acc": 0.082, "reasoning_bytes": 203.846, "flops_per_problem": 10614064398.336}, "x.add.rev9": {"acc": 0.0, "reasoning_bytes": 218.382, "flops_per_problem": 11445390065.664}, "x.add.rev10": {"acc": 0.0, "reasoning_bytes": 232.456, "flops_per_problem": 12167986249.728}, "x.add.sumd8": {"acc": 0.086, "reasoning_bytes": 202.212, "flops_per_problem": 10381678313.472}, "x.add.sumd9": {"acc": 0.044, "reasoning_bytes": 215.716, "flops_per_problem": 11071947128.832}, "x.add.sumd10": {"acc": 0.04, "reasoning_bytes": 228.668, "flops_per_problem": 11732626685.952}, "x.add.inc8": {"acc": 0.118, "reasoning_bytes": 197.948, "flops_per_problem": 10436440522.752}, "x.add.inc9": {"acc": 0.002, "reasoning_bytes": 206.826, "flops_per_problem": 10930801950.72}, "x.add.inc10": {"acc": 0.0, "reasoning_bytes": 219.774, "flops_per_problem": 11619745873.92}, "x.runmax.cmp12": {"acc": 0.858, "reasoning_bytes": 123.28, "flops_per_problem": 6813832052.736}, "x.runmax.cmp14": {"acc": 0.912, "reasoning_bytes": 135.872, "flops_per_problem": 7459672817.664}, "x.runmax.cmp16": {"acc": 0.978, "reasoning_bytes": 146.664, "flops_per_problem": 8022751985.664}, "x.runmax.runmax12": {"acc": 0.852, "reasoning_bytes": 126.3, "flops_per_problem": 6986156359.68}, "x.runmax.runmax14": {"acc": 0.794, "reasoning_bytes": 141.756, "flops_per_problem": 7758833467.392}, "x.runmax.runmax16": {"acc": 0.928, "reasoning_bytes": 153.5, "flops_per_problem": 8368548839.424}, "x.runmax.count12": {"acc": 0.954, "reasoning_bytes": 123.7, "flops_per_problem": 6920265056.256}, "x.runmax.count14": {"acc": 0.966, "reasoning_bytes": 138.194, "flops_per_problem": 7650280636.416}, "x.runmax.count16": {"acc": 0.996, "reasoning_bytes": 149.168, "flops_per_problem": 8221662461.952}, "x.runmax.find12": {"acc": 0.826, "reasoning_bytes": 124.792, "flops_per_problem": 6928567713.792}, "x.runmax.find14": {"acc": 0.894, "reasoning_bytes": 138.92, "flops_per_problem": 7638974889.984}, "x.runmax.find16": {"acc": 0.928, "reasoning_bytes": 149.336, "flops_per_problem": 8186596982.784}, "x.runmax.rev12": {"acc": 0.872, "reasoning_bytes": 118.732, "flops_per_problem": 6526153801.728}, "x.runmax.rev14": {"acc": 0.912, "reasoning_bytes": 132.794, "flops_per_problem": 7233822867.456}, "x.runmax.rev16": {"acc": 0.952, "reasoning_bytes": 143.122, "flops_per_problem": 7778000240.64}, "x.runmax.sumd12": {"acc": 0.868, "reasoning_bytes": 122.884, "flops_per_problem": 6763132846.08}, "x.runmax.sumd14": {"acc": 0.882, "reasoning_bytes": 136.902, "flops_per_problem": 7469212041.216}, "x.runmax.sumd16": {"acc": 0.944, "reasoning_bytes": 148.614, "flops_per_problem": 8063735316.48}, "x.runmax.inc12": {"acc": 0.88, "reasoning_bytes": 123.568, "flops_per_problem": 6768520740.864}, "x.runmax.inc14": {"acc": 0.926, "reasoning_bytes": 136.008, "flops_per_problem": 7409591894.016}, "x.runmax.inc16": {"acc": 0.774, "reasoning_bytes": 147.568, "flops_per_problem": 8002878603.264}, "x.count.cmp16": {"acc": 0.444, "reasoning_bytes": 133.918, "flops_per_problem": 7562926080.0}, "x.count.cmp18": {"acc": 0.516, "reasoning_bytes": 145.662, "flops_per_problem": 8161159053.312}, "x.count.cmp20": {"acc": 0.608, "reasoning_bytes": 156.33, "flops_per_problem": 8712932474.88}, "x.count.runmax16": {"acc": 0.264, "reasoning_bytes": 138.516, "flops_per_problem": 7809532674.048}, "x.count.runmax18": {"acc": 0.296, "reasoning_bytes": 149.67, "flops_per_problem": 8415803326.464}, "x.count.runmax20": {"acc": 0.368, "reasoning_bytes": 162.556, "flops_per_problem": 9031701528.576}, "x.count.count16": {"acc": 0.618, "reasoning_bytes": 136.738, "flops_per_problem": 7745937850.368}, "x.count.count18": {"acc": 0.656, "reasoning_bytes": 148.422, "flops_per_problem": 8350618632.192}, "x.count.count20": {"acc": 0.964, "reasoning_bytes": 160.734, "flops_per_problem": 8954504478.72}, "x.count.find16": {"acc": 0.554, "reasoning_bytes": 135.66, "flops_per_problem": 7649220722.688}, "x.count.find18": {"acc": 0.582, "reasoning_bytes": 147.446, "flops_per_problem": 8259642703.872}, "x.count.find20": {"acc": 0.656, "reasoning_bytes": 157.438, "flops_per_problem": 8797018963.968}, "x.count.rev16": {"acc": 0.33, "reasoning_bytes": 130.938, "flops_per_problem": 7310313308.16}, "x.count.rev18": {"acc": 0.318, "reasoning_bytes": 142.35, "flops_per_problem": 7918438809.6}, "x.count.rev20": {"acc": 0.412, "reasoning_bytes": 154.574, "flops_per_problem": 8518703284.224}, "x.count.sumd16": {"acc": 0.354, "reasoning_bytes": 134.446, "flops_per_problem": 7518939660.288}, "x.count.sumd18": {"acc": 0.358, "reasoning_bytes": 145.546, "flops_per_problem": 8119027482.624}, "x.count.sumd20": {"acc": 0.444, "reasoning_bytes": 157.836, "flops_per_problem": 8715935563.776}, "x.count.inc16": {"acc": 0.44, "reasoning_bytes": 133.744, "flops_per_problem": 7461792645.12}, "x.count.inc18": {"acc": 0.4, "reasoning_bytes": 145.678, "flops_per_problem": 8073274540.032}, "x.count.inc20": {"acc": 0.364, "reasoning_bytes": 158.522, "flops_per_problem": 8698270334.976}, "x.parity.cmp24": {"acc": 0.0, "reasoning_bytes": 208.888, "flops_per_problem": 11129094144.0}, "x.parity.cmp28": {"acc": 0.0, "reasoning_bytes": 235.366, "flops_per_problem": 12542312448.0}, "x.parity.cmp32": {"acc": 0.0, "reasoning_bytes": 264.58, "flops_per_problem": 13955530752.0}, "x.parity.runmax24": {"acc": 0.222, "reasoning_bytes": 212.368, "flops_per_problem": 11349909504.0}, "x.parity.runmax28": {"acc": 0.08, "reasoning_bytes": 239.982, "flops_per_problem": 12763127808.0}, "x.parity.runmax32": {"acc": 0.242, "reasoning_bytes": 268.502, "flops_per_problem": 14175374524.416}, "x.parity.count24": {"acc": 0.13, "reasoning_bytes": 210.78, "flops_per_problem": 11349909504.0}, "x.parity.count28": {"acc": 0.03, "reasoning_bytes": 238.54, "flops_per_problem": 12763127808.0}, "x.parity.count32": {"acc": 0.336, "reasoning_bytes": 267.58, "flops_per_problem": 14176346112.0}, "x.parity.find24": {"acc": 0.026, "reasoning_bytes": 210.724, "flops_per_problem": 11261583360.0}, "x.parity.find28": {"acc": 0.006, "reasoning_bytes": 238.308, "flops_per_problem": 12674801664.0}, "x.parity.find32": {"acc": 0.026, "reasoning_bytes": 266.528, "flops_per_problem": 14088019968.0}, "x.parity.rev24": {"acc": 0.126, "reasoning_bytes": 203.328, "flops_per_problem": 10847157092.352}, "x.parity.rev28": {"acc": 0.196, "reasoning_bytes": 217.93, "flops_per_problem": 11657637789.696}, "x.parity.rev32": {"acc": 0.128, "reasoning_bytes": 251.562, "flops_per_problem": 13336187830.272}, "x.parity.sumd24": {"acc": 0.034, "reasoning_bytes": 207.598, "flops_per_problem": 11084931072.0}, "x.parity.sumd28": {"acc": 0.114, "reasoning_bytes": 235.798, "flops_per_problem": 12498149376.0}, "x.parity.sumd32": {"acc": 0.036, "reasoning_bytes": 263.916, "flops_per_problem": 13911367680.0}, "x.parity.inc24": {"acc": 0.11, "reasoning_bytes": 208.93, "flops_per_problem": 11040768000.0}, "x.parity.inc28": {"acc": 0.016, "reasoning_bytes": 234.42, "flops_per_problem": 12453986304.0}, "x.parity.inc32": {"acc": 0.072, "reasoning_bytes": 264.46, "flops_per_problem": 13867204608.0}, "x.sumd.cmp8": {"acc": 0.95, "reasoning_bytes": 101.028, "flops_per_problem": 5609770057.728}, "x.sumd.cmp9": {"acc": 0.882, "reasoning_bytes": 110.188, "flops_per_problem": 6058643521.536}, "x.sumd.cmp10": {"acc": 0.848, "reasoning_bytes": 118.93, "flops_per_problem": 6488880168.96}, "x.sumd.runmax8": {"acc": 0.92, "reasoning_bytes": 109.86, "flops_per_problem": 5999376678.912}, "x.sumd.runmax9": {"acc": 0.814, "reasoning_bytes": 118.822, "flops_per_problem": 6439240876.032}, "x.sumd.runmax10": {"acc": 0.682, "reasoning_bytes": 127.786, "flops_per_problem": 6884846272.512}, "x.sumd.count8": {"acc": 0.856, "reasoning_bytes": 108.388, "flops_per_problem": 5978266730.496}, "x.sumd.count9": {"acc": 0.77, "reasoning_bytes": 117.254, "flops_per_problem": 6413184663.552}, "x.sumd.count10": {"acc": 0.598, "reasoning_bytes": 126.628, "flops_per_problem": 6873187221.504}, "x.sumd.find8": {"acc": 0.912, "reasoning_bytes": 107.202, "flops_per_problem": 5880931319.808}, "x.sumd.find9": {"acc": 0.854, "reasoning_bytes": 116.568, "flops_per_problem": 6340668899.328}, "x.sumd.find10": {"acc": 0.704, "reasoning_bytes": 125.602, "flops_per_problem": 6796961759.232}, "x.sumd.rev8": {"acc": 0.87, "reasoning_bytes": 101.154, "flops_per_problem": 5524182024.192}, "x.sumd.rev9": {"acc": 0.728, "reasoning_bytes": 110.15, "flops_per_problem": 5968197550.08}, "x.sumd.rev10": {"acc": 0.422, "reasoning_bytes": 119.626, "flops_per_problem": 6431026544.64}, "x.sumd.sumd8": {"acc": 0.79, "reasoning_bytes": 106.72, "flops_per_problem": 5784744148.992}, "x.sumd.sumd9": {"acc": 0.626, "reasoning_bytes": 115.426, "flops_per_problem": 6216394014.72}, "x.sumd.sumd10": {"acc": 0.434, "reasoning_bytes": 124.096, "flops_per_problem": 6640359505.92}, "x.sumd.inc8": {"acc": 0.276, "reasoning_bytes": 111.718, "flops_per_problem": 6000701571.072}, "x.sumd.inc9": {"acc": 0.104, "reasoning_bytes": 121.35, "flops_per_problem": 6484993818.624}, "x.sumd.inc10": {"acc": 0.03, "reasoning_bytes": 129.652, "flops_per_problem": 6926006255.616}, "x.inc.cmp8": {"acc": 0.958, "reasoning_bytes": 121.16, "flops_per_problem": 6719853035.52}, "x.inc.cmp9": {"acc": 0.946, "reasoning_bytes": 127.386, "flops_per_problem": 7083226791.936}, "x.inc.cmp10": {"acc": 0.914, "reasoning_bytes": 137.026, "flops_per_problem": 7597196623.872}, "x.inc.runmax8": {"acc": 0.968, "reasoning_bytes": 160.312, "flops_per_problem": 8182622306.304}, "x.inc.runmax9": {"acc": 0.934, "reasoning_bytes": 168.506, "flops_per_problem": 8589717504.0}, "x.inc.runmax10": {"acc": 0.924, "reasoning_bytes": 179.432, "flops_per_problem": 9116582952.96}, "x.inc.count8": {"acc": 0.774, "reasoning_bytes": 151.82, "flops_per_problem": 7853077463.04}, "x.inc.count9": {"acc": 0.68, "reasoning_bytes": 156.956, "flops_per_problem": 8124062072.832}, "x.inc.count10": {"acc": 0.534, "reasoning_bytes": 169.356, "flops_per_problem": 8715847237.632}, "x.inc.find8": {"acc": 0.702, "reasoning_bytes": 139.328, "flops_per_problem": 7257140969.472}, "x.inc.find9": {"acc": 0.656, "reasoning_bytes": 144.042, "flops_per_problem": 7509577089.024}, "x.inc.find10": {"acc": 0.56, "reasoning_bytes": 155.594, "flops_per_problem": 8064618577.92}, "x.inc.rev8": {"acc": 0.55, "reasoning_bytes": 166.652, "flops_per_problem": 8617275260.928}, "x.inc.rev9": {"acc": 0.044, "reasoning_bytes": 175.158, "flops_per_problem": 9072419880.96}, "x.inc.rev10": {"acc": 0.0, "reasoning_bytes": 193.366, "flops_per_problem": 9944993857.536}, "x.inc.sumd8": {"acc": 0.14, "reasoning_bytes": 171.012, "flops_per_problem": 8597490204.672}, "x.inc.sumd9": {"acc": 0.03, "reasoning_bytes": 174.602, "flops_per_problem": 8811327799.296}, "x.inc.sumd10": {"acc": 0.028, "reasoning_bytes": 192.104, "flops_per_problem": 9632849264.64}, "x.inc.inc8": {"acc": 0.344, "reasoning_bytes": 163.114, "flops_per_problem": 8471360471.04}, "x.inc.inc9": {"acc": 0.034, "reasoning_bytes": 173.022, "flops_per_problem": 8974112882.688}, "x.inc.inc10": {"acc": 0.002, "reasoning_bytes": 187.696, "flops_per_problem": 9710399619.072}} +{"step": 4000, "train_loss": 1.1628146171569824} +{"step": 4020, "train_loss": 1.1129871606826782} +{"step": 4040, "train_loss": 1.2311160564422607} +{"step": 4060, "train_loss": 1.0548095703125} +{"step": 4080, "train_loss": 1.232039451599121} +{"step": 4100, "train_loss": 1.2538259029388428} +{"step": 4120, "train_loss": 1.083634614944458} +{"step": 4140, "train_loss": 1.1776505708694458} +{"step": 4160, "train_loss": 1.2055801153182983} +{"step": 4180, "train_loss": 1.1164324283599854} +{"step": 4200, "train_loss": 1.1050525903701782} +{"step": 4220, "train_loss": 1.0833016633987427} +{"step": 4240, "train_loss": 1.1278676986694336} +{"step": 4260, "train_loss": 1.0742602348327637} +{"step": 4280, "train_loss": 1.170547604560852} +{"step": 4300, "train_loss": 1.1698875427246094} +{"step": 4320, "train_loss": 1.1121842861175537} +{"step": 4340, "train_loss": 1.1360273361206055} +{"step": 4360, "train_loss": 1.1420519351959229} +{"step": 4380, "train_loss": 1.0475575923919678} +{"step": 4400, "train_loss": 1.1825363636016846} +{"step": 4420, "train_loss": 1.1636312007904053} +{"step": 4440, "train_loss": 1.218172311782837} +{"step": 4460, "train_loss": 1.0456900596618652} +{"step": 4480, "train_loss": 1.0746533870697021} +{"step": 4500, "train_loss": 1.1082700490951538} +{"step": 4520, "train_loss": 1.1265225410461426} +{"step": 4540, "train_loss": 1.0570849180221558} +{"step": 4560, "train_loss": 1.1498236656188965} +{"step": 4580, "train_loss": 1.2006115913391113} +{"step": 4600, "train_loss": 1.1418075561523438} +{"step": 4620, "train_loss": 1.154589295387268} +{"step": 4640, "train_loss": 1.0149041414260864} +{"step": 4660, "train_loss": 1.1048377752304077} +{"step": 4680, "train_loss": 1.0243170261383057} +{"step": 4700, "train_loss": 1.0562411546707153} +{"step": 4720, "train_loss": 1.0931419134140015} +{"step": 4740, "train_loss": 1.0013668537139893} +{"step": 4760, "train_loss": 1.1322627067565918} +{"step": 4780, "train_loss": 1.1261661052703857} +{"step": 4800, "train_loss": 1.203604817390442} +{"step": 4820, "train_loss": 1.1196730136871338} +{"step": 4840, "train_loss": 1.1710079908370972} +{"step": 4860, "train_loss": 1.1527048349380493} +{"step": 4880, "train_loss": 1.1344549655914307} +{"step": 4900, "train_loss": 1.0467243194580078} +{"step": 4920, "train_loss": 1.1721652746200562} +{"step": 4940, "train_loss": 1.1336640119552612} +{"step": 4960, "train_loss": 1.1334571838378906} +{"step": 4980, "train_loss": 1.0792832374572754} +{"step": 5000, "train_loss": 1.102595567703247} +{"step": 5020, "train_loss": 1.1119318008422852} +{"step": 5040, "train_loss": 1.169561743736267} +{"step": 5060, "train_loss": 1.0830442905426025} +{"step": 5080, "train_loss": 1.0723987817764282} +{"step": 5100, "train_loss": 1.1466071605682373} +{"step": 5120, "train_loss": 1.0312472581863403} +{"step": 5140, "train_loss": 1.1243314743041992} +{"step": 5160, "train_loss": 1.109435796737671} +{"step": 5180, "train_loss": 1.103605031967163} +{"step": 5200, "train_loss": 1.0626450777053833} +{"step": 5220, "train_loss": 1.1781123876571655} +{"step": 5240, "train_loss": 1.0931525230407715} +{"step": 5260, "train_loss": 0.9611102342605591} +{"step": 5280, "train_loss": 1.1973366737365723} +{"step": 5300, "train_loss": 1.1093006134033203} +{"step": 5320, "train_loss": 1.1362569332122803} +{"step": 5340, "train_loss": 1.057030200958252} +{"step": 5360, "train_loss": 1.168677806854248} +{"step": 5380, "train_loss": 1.203881025314331} +{"step": 5400, "train_loss": 1.1757421493530273} +{"step": 5420, "train_loss": 1.0160183906555176} +{"step": 5440, "train_loss": 1.0644092559814453} +{"step": 5460, "train_loss": 1.115480899810791} +{"step": 5480, "train_loss": 1.0782488584518433} +{"step": 5500, "train_loss": 1.0573182106018066} +{"step": 5520, "train_loss": 1.0410469770431519} +{"step": 5540, "train_loss": 1.0696709156036377} +{"step": 5560, "train_loss": 1.089282512664795} +{"step": 5580, "train_loss": 1.2018405199050903} +{"step": 5600, "train_loss": 1.1326453685760498} +{"step": 5620, "train_loss": 1.0417442321777344} +{"step": 5640, "train_loss": 1.0478113889694214} +{"step": 5660, "train_loss": 1.053248405456543} +{"step": 5680, "train_loss": 1.1372665166854858} +{"step": 5700, "train_loss": 1.1212868690490723} +{"step": 5720, "train_loss": 1.0820605754852295} +{"step": 5740, "train_loss": 1.0973809957504272} +{"step": 5760, "train_loss": 1.0571849346160889} +{"step": 5780, "train_loss": 1.0422332286834717} +{"step": 5800, "train_loss": 1.0835824012756348} +{"step": 5820, "train_loss": 1.0527540445327759} +{"step": 5840, "train_loss": 1.0884947776794434} +{"step": 5860, "train_loss": 1.1198500394821167} +{"step": 5880, "train_loss": 1.0659799575805664} +{"step": 5900, "train_loss": 0.995494544506073} +{"step": 5920, "train_loss": 0.9673564434051514} +{"step": 5940, "train_loss": 0.9763011932373047} +{"step": 5960, "train_loss": 1.0735628604888916} +{"step": 5980, "train_loss": 1.0618749856948853} +{"step": 6000, "train_loss": 1.1288844347000122} +{"step": 6020, "train_loss": 1.0425945520401} +{"step": 6040, "train_loss": 1.0835586786270142} +{"step": 6060, "train_loss": 1.1460002660751343} +{"step": 6080, "train_loss": 1.0683221817016602} +{"step": 6100, "train_loss": 1.1468236446380615} +{"step": 6120, "train_loss": 1.191359519958496} +{"step": 6140, "train_loss": 1.1624270677566528} +{"step": 6160, "train_loss": 1.111738681793213} +{"step": 6180, "train_loss": 1.1356909275054932} +{"step": 6200, "train_loss": 1.1029099225997925} +{"step": 6220, "train_loss": 1.1361998319625854} +{"step": 6240, "train_loss": 1.1422035694122314} +{"step": 6260, "train_loss": 1.039940357208252} +{"step": 6280, "train_loss": 1.0310546159744263} +{"step": 6300, "train_loss": 1.0174250602722168} +{"step": 6320, "train_loss": 1.1068178415298462} +{"step": 6340, "train_loss": 1.0349489450454712} +{"step": 6360, "train_loss": 1.036027431488037} +{"step": 6380, "train_loss": 1.0722508430480957} +{"step": 6400, "train_loss": 1.0795209407806396} +{"step": 6420, "train_loss": 1.0572357177734375} +{"step": 6440, "train_loss": 1.1100499629974365} +{"step": 6460, "train_loss": 1.0668846368789673} +{"step": 6480, "train_loss": 1.043851613998413} +{"step": 6500, "train_loss": 1.0442347526550293} +{"step": 6520, "train_loss": 1.060969352722168} +{"step": 6540, "train_loss": 1.0591762065887451} +{"step": 6560, "train_loss": 1.1151453256607056} +{"step": 6580, "train_loss": 1.0644011497497559} +{"step": 6600, "train_loss": 1.203298568725586} +{"step": 6620, "train_loss": 1.101731538772583} +{"step": 6640, "train_loss": 1.1620843410491943} +{"step": 6660, "train_loss": 1.072969675064087} +{"step": 6680, "train_loss": 1.0866210460662842} +{"step": 6700, "train_loss": 1.0846316814422607} +{"step": 6720, "train_loss": 1.0655169486999512} +{"step": 6740, "train_loss": 1.075329065322876} +{"step": 6760, "train_loss": 1.1219462156295776} +{"step": 6780, "train_loss": 1.0448143482208252} +{"step": 6800, "train_loss": 1.1021604537963867} +{"step": 6820, "train_loss": 1.0578336715698242} +{"step": 6840, "train_loss": 0.9666056632995605} +{"step": 6860, "train_loss": 1.0947914123535156} +{"step": 6880, "train_loss": 1.0572433471679688} +{"step": 6900, "train_loss": 1.0022971630096436} +{"step": 6920, "train_loss": 1.0561776161193848} +{"step": 6940, "train_loss": 1.0970847606658936} +{"step": 6960, "train_loss": 1.0370641946792603} +{"step": 6980, "train_loss": 1.0214217901229858} +{"step": 7000, "train_loss": 0.9772053956985474} +{"step": 7020, "train_loss": 1.0219111442565918} +{"step": 7040, "train_loss": 1.0986931324005127} +{"step": 7060, "train_loss": 1.0054361820220947} +{"step": 7080, "train_loss": 1.1308801174163818} +{"step": 7100, "train_loss": 1.0446536540985107} +{"step": 7120, "train_loss": 1.0746325254440308} +{"step": 7140, "train_loss": 1.1209379434585571} +{"step": 7160, "train_loss": 1.140160083770752} +{"step": 7180, "train_loss": 1.0544047355651855} +{"step": 7200, "train_loss": 1.121814250946045} +{"step": 7220, "train_loss": 1.0183241367340088} +{"step": 7240, "train_loss": 1.0163822174072266} +{"step": 7260, "train_loss": 1.0903666019439697} +{"step": 7280, "train_loss": 1.0337588787078857} +{"step": 7300, "train_loss": 1.0622215270996094} +{"step": 7320, "train_loss": 1.0313736200332642} +{"step": 7340, "train_loss": 1.0150973796844482} +{"step": 7360, "train_loss": 1.027725338935852} +{"step": 7380, "train_loss": 1.0497528314590454} +{"step": 7400, "train_loss": 1.0495915412902832} +{"step": 7420, "train_loss": 1.0718612670898438} +{"step": 7440, "train_loss": 1.0136635303497314} +{"step": 7460, "train_loss": 1.077182412147522} +{"step": 7480, "train_loss": 1.054754614830017} +{"step": 7500, "train_loss": 1.00557279586792} +{"step": 7520, "train_loss": 1.052679419517517} +{"step": 7540, "train_loss": 0.9942553639411926} +{"step": 7560, "train_loss": 1.0684359073638916} +{"step": 7580, "train_loss": 1.008618950843811} +{"step": 7600, "train_loss": 1.0473732948303223} +{"step": 7620, "train_loss": 1.0134133100509644} +{"step": 7640, "train_loss": 1.039269208908081} +{"step": 7660, "train_loss": 1.0707743167877197} +{"step": 7680, "train_loss": 1.1156971454620361} +{"step": 7700, "train_loss": 1.0063873529434204} +{"step": 7720, "train_loss": 1.0061872005462646} +{"step": 7740, "train_loss": 1.0296778678894043} +{"step": 7760, "train_loss": 1.0545852184295654} +{"step": 7780, "train_loss": 1.0682092905044556} +{"step": 7800, "train_loss": 0.9638962149620056} +{"step": 7820, "train_loss": 1.0438730716705322} +{"step": 7840, "train_loss": 1.0468976497650146} +{"step": 7860, "train_loss": 1.0276126861572266} +{"step": 7880, "train_loss": 1.0253139734268188} +{"step": 7900, "train_loss": 1.13986074924469} +{"step": 7920, "train_loss": 1.005910038948059} +{"step": 7940, "train_loss": 1.0253369808197021} +{"step": 7960, "train_loss": 0.9813728928565979} +{"step": 7980, "train_loss": 1.072350025177002} +{"step": 8000, "tokens": 262144000, "val_loss": 1.8880574762821198, "val_bpb": 0.6806140463355776, "lr": 0.00020868460938776225, "elapsed": 581.5149762630463, "peak_mem_gb": 2.254330368, "add8": {"acc": 0.984, "reasoning_bytes": 79.0, "flops_per_problem": 4928775487.488}, "add9": {"acc": 0.88, "reasoning_bytes": 89.0, "flops_per_problem": 5504220315.648}, "add10": {"acc": 0.066, "reasoning_bytes": 96.2, "flops_per_problem": 5945674383.36}, "reach14": {"acc": 0.56, "reasoning_bytes": 9.504, "flops_per_problem": 4703543820.288}, "reach16": {"acc": 0.556, "reasoning_bytes": 11.246, "flops_per_problem": 5310432755.712}, "reach18": {"acc": 0.536, "reasoning_bytes": 11.306, "flops_per_problem": 5843039404.032}, "parity24": {"acc": 0.378, "reasoning_bytes": 136.088, "flops_per_problem": 7555771662.336}, "parity28": {"acc": 0.278, "reasoning_bytes": 155.528, "flops_per_problem": 8590954070.016}, "parity32": {"acc": 0.602, "reasoning_bytes": 176.676, "flops_per_problem": 9701567004.672}, "cmp8": {"acc": 0.996, "reasoning_bytes": 5.78, "flops_per_problem": 1359339356.16}, "cmp9": {"acc": 0.996, "reasoning_bytes": 5.72, "flops_per_problem": 1445015715.84}, "cmp10": {"acc": 0.996, "reasoning_bytes": 5.768, "flops_per_problem": 1535461687.296}, "runmax12": {"acc": 1.0, "reasoning_bytes": 71.0, "flops_per_problem": 4151328768.0}, "runmax14": {"acc": 0.988, "reasoning_bytes": 82.016, "flops_per_problem": 4726155313.152}, "runmax16": {"acc": 0.988, "reasoning_bytes": 89.056, "flops_per_problem": 5125389484.032}, "count16": {"acc": 0.988, "reasoning_bytes": 79.05, "flops_per_problem": 4727656857.6}, "count18": {"acc": 0.964, "reasoning_bytes": 88.128, "flops_per_problem": 5216895369.216}, "count20": {"acc": 0.82, "reasoning_bytes": 95.528, "flops_per_problem": 5631939919.872}, "brackets16": {"acc": 0.93, "reasoning_bytes": 49.874, "flops_per_problem": 3483229814.784}, "brackets18": {"acc": 0.938, "reasoning_bytes": 52.102, "flops_per_problem": 3670039609.344}, "brackets20": {"acc": 0.932, "reasoning_bytes": 58.42, "flops_per_problem": 4037388042.24}, "sumd8": {"acc": 0.99, "reasoning_bytes": 58.566, "flops_per_problem": 3381301444.608}, "sumd9": {"acc": 0.91, "reasoning_bytes": 66.268, "flops_per_problem": 3765255192.576}, "sumd10": {"acc": 0.564, "reasoning_bytes": 72.638, "flops_per_problem": 4089677119.488}, "inc8": {"acc": 0.998, "reasoning_bytes": 63.0, "flops_per_problem": 3798024192.0}, "inc9": {"acc": 0.848, "reasoning_bytes": 70.072, "flops_per_problem": 4193460338.688}, "inc10": {"acc": 0.014, "reasoning_bytes": 71.176, "flops_per_problem": 4292650598.4}, "add.cmp6": {"acc": 1.0, "reasoning_bytes": 92.806, "flops_per_problem": 5540169056.256}, "add.cmp7": {"acc": 1.0, "reasoning_bytes": 106.1, "flops_per_problem": 6257024040.96}, "add.cmp8": {"acc": 0.988, "reasoning_bytes": 119.822, "flops_per_problem": 6997020475.392}, "add.runmax6": {"acc": 1.0, "reasoning_bytes": 120.832, "flops_per_problem": 6572878331.904}, "add.runmax7": {"acc": 0.996, "reasoning_bytes": 138.868, "flops_per_problem": 7457641316.352}, "add.runmax8": {"acc": 0.902, "reasoning_bytes": 156.866, "flops_per_problem": 8340637777.92}, "add.count6": {"acc": 0.982, "reasoning_bytes": 115.648, "flops_per_problem": 6388100038.656}, "add.count7": {"acc": 0.956, "reasoning_bytes": 132.102, "flops_per_problem": 7203085369.344}, "add.count8": {"acc": 0.396, "reasoning_bytes": 144.078, "flops_per_problem": 7820308463.616}, "add.find6": {"acc": 0.956, "reasoning_bytes": 107.696, "flops_per_problem": 5992752218.112}, "add.find7": {"acc": 0.87, "reasoning_bytes": 121.236, "flops_per_problem": 6678958030.848}, "add.find8": {"acc": 0.542, "reasoning_bytes": 134.954, "flops_per_problem": 7373113196.544}, "add.rev6": {"acc": 0.966, "reasoning_bytes": 123.256, "flops_per_problem": 6795018584.064}, "add.rev7": {"acc": 0.92, "reasoning_bytes": 145.932, "flops_per_problem": 7928419663.872}, "add.rev8": {"acc": 0.368, "reasoning_bytes": 169.726, "flops_per_problem": 9110046818.304}, "add.sumd6": {"acc": 0.994, "reasoning_bytes": 126.462, "flops_per_problem": 6777088376.832}, "add.sumd7": {"acc": 0.952, "reasoning_bytes": 146.326, "flops_per_problem": 7742846435.328}, "add.sumd8": {"acc": 0.612, "reasoning_bytes": 166.414, "flops_per_problem": 8718320369.664}, "add.inc6": {"acc": 0.998, "reasoning_bytes": 131.04, "flops_per_problem": 7138695610.368}, "add.inc7": {"acc": 0.962, "reasoning_bytes": 151.064, "flops_per_problem": 8155771158.528}, "add.inc8": {"acc": 0.664, "reasoning_bytes": 170.924, "flops_per_problem": 9166663876.608}, "runmax.cmp12": {"acc": 1.0, "reasoning_bytes": 87.0, "flops_per_problem": 5122916352.0}, "runmax.cmp14": {"acc": 0.994, "reasoning_bytes": 96.972, "flops_per_problem": 5651636649.984}, "runmax.cmp16": {"acc": 0.972, "reasoning_bytes": 104.486, "flops_per_problem": 6072775704.576}, "runmax.runmax12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5211242496.0}, "runmax.runmax14": {"acc": 0.972, "reasoning_bytes": 96.076, "flops_per_problem": 5656229609.472}, "runmax.runmax16": {"acc": 0.972, "reasoning_bytes": 105.948, "flops_per_problem": 6180710252.544}, "runmax.count12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5255405568.0}, "runmax.count14": {"acc": 0.972, "reasoning_bytes": 96.256, "flops_per_problem": 5708342034.432}, "runmax.count16": {"acc": 0.98, "reasoning_bytes": 105.958, "flops_per_problem": 6225049976.832}, "runmax.find12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5211242496.0}, "runmax.find14": {"acc": 0.994, "reasoning_bytes": 96.712, "flops_per_problem": 5684317323.264}, "runmax.find16": {"acc": 0.98, "reasoning_bytes": 105.944, "flops_per_problem": 6180886904.832}, "runmax.rev12": {"acc": 1.0, "reasoning_bytes": 83.0, "flops_per_problem": 4857937920.0}, "runmax.rev14": {"acc": 0.972, "reasoning_bytes": 91.388, "flops_per_problem": 5316703911.936}, "runmax.rev16": {"acc": 0.966, "reasoning_bytes": 100.924, "flops_per_problem": 5826434088.96}, "runmax.sumd12": {"acc": 1.0, "reasoning_bytes": 86.0, "flops_per_problem": 5034590208.0}, "runmax.sumd14": {"acc": 0.974, "reasoning_bytes": 94.18, "flops_per_problem": 5484170280.96}, "runmax.sumd16": {"acc": 0.97, "reasoning_bytes": 103.862, "flops_per_problem": 6000259940.352}, "runmax.inc12": {"acc": 0.854, "reasoning_bytes": 87.0, "flops_per_problem": 5059321528.32}, "runmax.inc14": {"acc": 0.21, "reasoning_bytes": 96.168, "flops_per_problem": 5527980048.384}, "runmax.inc16": {"acc": 0.158, "reasoning_bytes": 104.714, "flops_per_problem": 5993723805.696}, "count.cmp16": {"acc": 0.996, "reasoning_bytes": 94.98, "flops_per_problem": 5696153026.56}, "count.cmp18": {"acc": 0.988, "reasoning_bytes": 104.32, "flops_per_problem": 6197050589.184}, "count.cmp20": {"acc": 0.958, "reasoning_bytes": 111.18, "flops_per_problem": 6588335407.104}, "count.runmax16": {"acc": 0.988, "reasoning_bytes": 96.046, "flops_per_problem": 5787305607.168}, "count.runmax18": {"acc": 0.966, "reasoning_bytes": 105.57, "flops_per_problem": 6296329175.04}, "count.runmax20": {"acc": 0.826, "reasoning_bytes": 113.126, "flops_per_problem": 6718263164.928}, "count.count16": {"acc": 0.998, "reasoning_bytes": 96.02, "flops_per_problem": 5830408765.44}, "count.count18": {"acc": 0.996, "reasoning_bytes": 105.63, "flops_per_problem": 6343142031.36}, "count.count20": {"acc": 0.992, "reasoning_bytes": 112.994, "flops_per_problem": 6756685037.568}, "count.find16": {"acc": 0.996, "reasoning_bytes": 96.01, "flops_per_problem": 5785804062.72}, "count.find18": {"acc": 0.984, "reasoning_bytes": 105.23, "flops_per_problem": 6281313730.56}, "count.find20": {"acc": 0.956, "reasoning_bytes": 112.57, "flops_per_problem": 6693796823.04}, "count.rev16": {"acc": 0.988, "reasoning_bytes": 91.046, "flops_per_problem": 5434001031.168}, "count.rev18": {"acc": 0.966, "reasoning_bytes": 100.38, "flops_per_problem": 5934633615.36}, "count.rev20": {"acc": 0.812, "reasoning_bytes": 107.636, "flops_per_problem": 6343318683.648}, "count.sumd16": {"acc": 0.986, "reasoning_bytes": 94.036, "flops_per_problem": 5610211688.448}, "count.sumd18": {"acc": 0.964, "reasoning_bytes": 103.42, "flops_per_problem": 6113052426.24}, "count.sumd20": {"acc": 0.82, "reasoning_bytes": 110.906, "flops_per_problem": 6531895001.088}, "count.inc16": {"acc": 0.988, "reasoning_bytes": 95.05, "flops_per_problem": 5610918297.6}, "count.inc18": {"acc": 0.97, "reasoning_bytes": 104.606, "flops_per_problem": 6121266757.632}, "count.inc20": {"acc": 0.826, "reasoning_bytes": 112.224, "flops_per_problem": 6546203836.416}, "parity.cmp16": {"acc": 0.994, "reasoning_bytes": 110.724, "flops_per_problem": 6347293360.128}, "parity.cmp20": {"acc": 0.978, "reasoning_bytes": 129.864, "flops_per_problem": 7369226846.208}, "parity.cmp24": {"acc": 0.916, "reasoning_bytes": 147.156, "flops_per_problem": 8309546975.232}, "parity.runmax16": {"acc": 0.996, "reasoning_bytes": 111.976, "flops_per_problem": 6446748598.272}, "parity.runmax20": {"acc": 0.728, "reasoning_bytes": 132.04, "flops_per_problem": 7509488762.88}, "parity.runmax24": {"acc": 0.282, "reasoning_bytes": 148.958, "flops_per_problem": 8433291902.976}, "parity.count16": {"acc": 0.998, "reasoning_bytes": 111.988, "flops_per_problem": 6491441627.136}, "parity.count20": {"acc": 0.966, "reasoning_bytes": 131.344, "flops_per_problem": 7522914336.768}, "parity.count24": {"acc": 0.996, "reasoning_bytes": 148.602, "flops_per_problem": 8461732921.344}, "parity.find16": {"acc": 1.0, "reasoning_bytes": 111.88, "flops_per_problem": 6442508943.36}, "parity.find20": {"acc": 0.904, "reasoning_bytes": 131.032, "flops_per_problem": 7464972386.304}, "parity.find24": {"acc": 0.836, "reasoning_bytes": 148.3, "flops_per_problem": 8404232601.6}, "parity.rev16": {"acc": 0.97, "reasoning_bytes": 106.82, "flops_per_problem": 6086554583.04}, "parity.rev20": {"acc": 0.716, "reasoning_bytes": 126.944, "flops_per_problem": 7151944531.968}, "parity.rev24": {"acc": 0.258, "reasoning_bytes": 143.824, "flops_per_problem": 8074069475.328}, "parity.sumd16": {"acc": 0.978, "reasoning_bytes": 109.868, "flops_per_problem": 6265326698.496}, "parity.sumd20": {"acc": 0.716, "reasoning_bytes": 129.86, "flops_per_problem": 7324887121.92}, "parity.sumd24": {"acc": 0.276, "reasoning_bytes": 147.032, "flops_per_problem": 8259907682.304}, "parity.inc16": {"acc": 0.984, "reasoning_bytes": 110.904, "flops_per_problem": 6266916569.088}, "parity.inc20": {"acc": 0.742, "reasoning_bytes": 131.028, "flops_per_problem": 7332306518.016}, "parity.inc24": {"acc": 0.228, "reasoning_bytes": 147.924, "flops_per_problem": 8255138070.528}, "sumd.cmp6": {"acc": 0.998, "reasoning_bytes": 62.064, "flops_per_problem": 3712347832.32}, "sumd.cmp7": {"acc": 0.988, "reasoning_bytes": 69.492, "flops_per_problem": 4084730855.424}, "sumd.cmp8": {"acc": 0.964, "reasoning_bytes": 74.29, "flops_per_problem": 4340700020.736}, "sumd.runmax6": {"acc": 1.0, "reasoning_bytes": 67.542, "flops_per_problem": 3954449793.024}, "sumd.runmax7": {"acc": 0.996, "reasoning_bytes": 75.464, "flops_per_problem": 4348472721.408}, "sumd.runmax8": {"acc": 0.904, "reasoning_bytes": 82.734, "flops_per_problem": 4713701326.848}, "sumd.count6": {"acc": 1.0, "reasoning_bytes": 66.428, "flops_per_problem": 3949415202.816}, "sumd.count7": {"acc": 0.994, "reasoning_bytes": 74.412, "flops_per_problem": 4346176241.664}, "sumd.count8": {"acc": 0.932, "reasoning_bytes": 81.834, "flops_per_problem": 4718117634.048}, "sumd.find6": {"acc": 1.0, "reasoning_bytes": 66.24, "flops_per_problem": 3896949473.28}, "sumd.find7": {"acc": 0.996, "reasoning_bytes": 74.294, "flops_per_problem": 4296801927.168}, "sumd.find8": {"acc": 0.928, "reasoning_bytes": 80.348, "flops_per_problem": 4608328237.056}, "sumd.rev6": {"acc": 0.956, "reasoning_bytes": 61.544, "flops_per_problem": 3601145217.024}, "sumd.rev7": {"acc": 0.968, "reasoning_bytes": 69.48, "flops_per_problem": 3995874754.56}, "sumd.rev8": {"acc": 0.942, "reasoning_bytes": 77.28, "flops_per_problem": 4384421462.016}, "sumd.sumd6": {"acc": 1.0, "reasoning_bytes": 65.744, "flops_per_problem": 3795639386.112}, "sumd.sumd7": {"acc": 0.994, "reasoning_bytes": 73.736, "flops_per_problem": 4195050209.28}, "sumd.sumd8": {"acc": 0.942, "reasoning_bytes": 81.474, "flops_per_problem": 4581565415.424}, "sumd.inc6": {"acc": 0.998, "reasoning_bytes": 68.538, "flops_per_problem": 3910021742.592}, "sumd.inc7": {"acc": 0.888, "reasoning_bytes": 76.464, "flops_per_problem": 4304309649.408}, "sumd.inc8": {"acc": 0.632, "reasoning_bytes": 84.09, "flops_per_problem": 4685436960.768}, "inc.cmp6": {"acc": 0.994, "reasoning_bytes": 78.662, "flops_per_problem": 4578032369.664}, "inc.cmp7": {"acc": 0.99, "reasoning_bytes": 89.876, "flops_per_problem": 5161603203.072}, "inc.cmp8": {"acc": 0.966, "reasoning_bytes": 100.824, "flops_per_problem": 5733426659.328}, "inc.runmax6": {"acc": 1.0, "reasoning_bytes": 104.0, "flops_per_problem": 5520384000.0}, "inc.runmax7": {"acc": 1.0, "reasoning_bytes": 119.968, "flops_per_problem": 6269743005.696}, "inc.runmax8": {"acc": 0.998, "reasoning_bytes": 134.992, "flops_per_problem": 6977412071.424}, "inc.count6": {"acc": 0.972, "reasoning_bytes": 98.998, "flops_per_problem": 5343643385.856}, "inc.count7": {"acc": 0.96, "reasoning_bytes": 114.0, "flops_per_problem": 6050340864.0}, "inc.count8": {"acc": 0.784, "reasoning_bytes": 123.652, "flops_per_problem": 6520765906.944}, "inc.find6": {"acc": 0.93, "reasoning_bytes": 91.88, "flops_per_problem": 4985127567.36}, "inc.find7": {"acc": 0.9, "reasoning_bytes": 102.544, "flops_per_problem": 5500245639.168}, "inc.find8": {"acc": 0.68, "reasoning_bytes": 112.848, "flops_per_problem": 5999465005.056}, "inc.rev6": {"acc": 0.964, "reasoning_bytes": 104.0, "flops_per_problem": 5608533491.712}, "inc.rev7": {"acc": 0.88, "reasoning_bytes": 123.99, "flops_per_problem": 6579502792.704}, "inc.rev8": {"acc": 0.688, "reasoning_bytes": 144.012, "flops_per_problem": 7548705570.816}, "inc.sumd6": {"acc": 0.998, "reasoning_bytes": 109.558, "flops_per_problem": 5721590956.032}, "inc.sumd7": {"acc": 0.98, "reasoning_bytes": 127.446, "flops_per_problem": 6555743059.968}, "inc.sumd8": {"acc": 0.848, "reasoning_bytes": 144.504, "flops_per_problem": 7353239814.144}, "inc.inc6": {"acc": 0.992, "reasoning_bytes": 113.0, "flops_per_problem": 6006089465.856}, "inc.inc7": {"acc": 0.968, "reasoning_bytes": 131.0, "flops_per_problem": 6889174253.568}, "inc.inc8": {"acc": 0.678, "reasoning_bytes": 148.188, "flops_per_problem": 7735338713.088}, "x.add.cmp8": {"acc": 0.762, "reasoning_bytes": 150.162, "flops_per_problem": 8429052248.064}, "x.add.cmp9": {"acc": 0.658, "reasoning_bytes": 156.054, "flops_per_problem": 8846569930.752}, "x.add.cmp10": {"acc": 0.548, "reasoning_bytes": 148.672, "flops_per_problem": 8694648963.072}, "x.add.runmax8": {"acc": 0.984, "reasoning_bytes": 189.894, "flops_per_problem": 9887758516.224}, "x.add.runmax9": {"acc": 0.96, "reasoning_bytes": 209.198, "flops_per_problem": 10828696928.256}, "x.add.runmax10": {"acc": 0.748, "reasoning_bytes": 225.228, "flops_per_problem": 11625045442.56}, "x.add.count8": {"acc": 0.892, "reasoning_bytes": 181.246, "flops_per_problem": 9550087667.712}, "x.add.count9": {"acc": 0.724, "reasoning_bytes": 200.068, "flops_per_problem": 10469562826.752}, "x.add.count10": {"acc": 0.39, "reasoning_bytes": 213.982, "flops_per_problem": 11172373954.56}, "x.add.find8": {"acc": 0.836, "reasoning_bytes": 165.736, "flops_per_problem": 8820955348.992}, "x.add.find9": {"acc": 0.766, "reasoning_bytes": 182.764, "flops_per_problem": 9661290283.008}, "x.add.find10": {"acc": 0.552, "reasoning_bytes": 199.756, "flops_per_problem": 10500830281.728}, "x.add.rev8": {"acc": 0.634, "reasoning_bytes": 199.864, "flops_per_problem": 10529712930.816}, "x.add.rev9": {"acc": 0.242, "reasoning_bytes": 226.212, "flops_per_problem": 11827842269.184}, "x.add.rev10": {"acc": 0.01, "reasoning_bytes": 246.318, "flops_per_problem": 12823807868.928}, "x.add.sumd8": {"acc": 0.826, "reasoning_bytes": 197.512, "flops_per_problem": 10179323117.568}, "x.add.sumd9": {"acc": 0.574, "reasoning_bytes": 219.39, "flops_per_problem": 11233937276.928}, "x.add.sumd10": {"acc": 0.116, "reasoning_bytes": 239.236, "flops_per_problem": 12198282117.12}, "x.add.inc8": {"acc": 0.724, "reasoning_bytes": 201.226, "flops_per_problem": 10597282430.976}, "x.add.inc9": {"acc": 0.34, "reasoning_bytes": 222.06, "flops_per_problem": 11650395045.888}, "x.add.inc10": {"acc": 0.01, "reasoning_bytes": 237.65, "flops_per_problem": 12454516260.864}, "x.runmax.cmp12": {"acc": 1.0, "reasoning_bytes": 118.066, "flops_per_problem": 6583212490.752}, "x.runmax.cmp14": {"acc": 0.996, "reasoning_bytes": 131.458, "flops_per_problem": 7262970494.976}, "x.runmax.cmp16": {"acc": 0.998, "reasoning_bytes": 141.276, "flops_per_problem": 7784889679.872}, "x.runmax.runmax12": {"acc": 1.0, "reasoning_bytes": 122.056, "flops_per_problem": 6803586220.032}, "x.runmax.runmax14": {"acc": 0.988, "reasoning_bytes": 135.146, "flops_per_problem": 7470006976.512}, "x.runmax.runmax16": {"acc": 0.994, "reasoning_bytes": 144.46, "flops_per_problem": 7969667973.12}, "x.runmax.count12": {"acc": 1.0, "reasoning_bytes": 121.072, "flops_per_problem": 6804292829.184}, "x.runmax.count14": {"acc": 0.992, "reasoning_bytes": 134.238, "flops_per_problem": 7474069979.136}, "x.runmax.count16": {"acc": 0.996, "reasoning_bytes": 143.82, "flops_per_problem": 7985566679.04}, "x.runmax.find12": {"acc": 1.0, "reasoning_bytes": 120.122, "flops_per_problem": 6718174838.784}, "x.runmax.find14": {"acc": 0.998, "reasoning_bytes": 133.694, "flops_per_problem": 7405882195.968}, "x.runmax.find16": {"acc": 0.998, "reasoning_bytes": 144.2, "flops_per_problem": 7958185574.4}, "x.runmax.rev12": {"acc": 1.0, "reasoning_bytes": 114.056, "flops_per_problem": 6317792428.032}, "x.runmax.rev14": {"acc": 0.994, "reasoning_bytes": 127.458, "flops_per_problem": 6997992062.976}, "x.runmax.rev16": {"acc": 0.994, "reasoning_bytes": 136.562, "flops_per_problem": 7488378814.464}, "x.runmax.sumd12": {"acc": 1.0, "reasoning_bytes": 118.04, "flops_per_problem": 6537901178.88}, "x.runmax.sumd14": {"acc": 0.996, "reasoning_bytes": 131.362, "flops_per_problem": 7214567768.064}, "x.runmax.sumd16": {"acc": 0.992, "reasoning_bytes": 140.676, "flops_per_problem": 7714228764.672}, "x.runmax.inc12": {"acc": 0.974, "reasoning_bytes": 118.118, "flops_per_problem": 6527213715.456}, "x.runmax.inc14": {"acc": 0.326, "reasoning_bytes": 131.274, "flops_per_problem": 7171376283.648}, "x.runmax.inc16": {"acc": 0.158, "reasoning_bytes": 140.74, "flops_per_problem": 7672803803.136}, "x.count.cmp16": {"acc": 0.994, "reasoning_bytes": 129.814, "flops_per_problem": 7322855620.608}, "x.count.cmp18": {"acc": 0.99, "reasoning_bytes": 141.19, "flops_per_problem": 7913669197.824}, "x.count.cmp20": {"acc": 0.948, "reasoning_bytes": 150.264, "flops_per_problem": 8403084361.728}, "x.count.runmax16": {"acc": 0.984, "reasoning_bytes": 134.0, "flops_per_problem": 7551885312.0}, "x.count.runmax18": {"acc": 0.958, "reasoning_bytes": 145.46, "flops_per_problem": 8146320261.12}, "x.count.runmax20": {"acc": 0.878, "reasoning_bytes": 155.088, "flops_per_problem": 8659848462.336}, "x.count.count16": {"acc": 0.996, "reasoning_bytes": 132.93, "flops_per_problem": 7548793896.96}, "x.count.count18": {"acc": 0.998, "reasoning_bytes": 144.126, "flops_per_problem": 8131569795.072}, "x.count.count20": {"acc": 0.996, "reasoning_bytes": 153.844, "flops_per_problem": 8648896020.48}, "x.count.find16": {"acc": 0.898, "reasoning_bytes": 131.996, "flops_per_problem": 7463382515.712}, "x.count.find18": {"acc": 0.892, "reasoning_bytes": 143.38, "flops_per_problem": 8054461071.36}, "x.count.find20": {"acc": 0.858, "reasoning_bytes": 152.828, "flops_per_problem": 8559951593.472}, "x.count.rev16": {"acc": 0.984, "reasoning_bytes": 125.978, "flops_per_problem": 7065119932.416}, "x.count.rev18": {"acc": 0.966, "reasoning_bytes": 137.41, "flops_per_problem": 7658318315.52}, "x.count.rev20": {"acc": 0.868, "reasoning_bytes": 146.992, "flops_per_problem": 8169815015.424}, "x.count.sumd16": {"acc": 0.978, "reasoning_bytes": 129.974, "flops_per_problem": 7285758640.128}, "x.count.sumd18": {"acc": 0.96, "reasoning_bytes": 141.304, "flops_per_problem": 7874452389.888}, "x.count.sumd20": {"acc": 0.858, "reasoning_bytes": 150.556, "flops_per_problem": 8371286949.888}, "x.count.inc16": {"acc": 0.98, "reasoning_bytes": 129.972, "flops_per_problem": 7241507241.984}, "x.count.inc18": {"acc": 0.958, "reasoning_bytes": 141.41, "flops_per_problem": 7834970603.52}, "x.count.inc20": {"acc": 0.86, "reasoning_bytes": 151.08, "flops_per_problem": 8350353653.76}, "x.parity.cmp24": {"acc": 0.97, "reasoning_bytes": 198.892, "flops_per_problem": 10682693812.224}, "x.parity.cmp28": {"acc": 0.934, "reasoning_bytes": 219.604, "flops_per_problem": 11774051647.488}, "x.parity.cmp32": {"acc": 0.916, "reasoning_bytes": 241.078, "flops_per_problem": 12898796765.184}, "x.parity.runmax24": {"acc": 0.748, "reasoning_bytes": 203.544, "flops_per_problem": 10932303495.168}, "x.parity.runmax28": {"acc": 0.392, "reasoning_bytes": 225.252, "flops_per_problem": 12067647750.144}, "x.parity.runmax32": {"acc": 0.396, "reasoning_bytes": 245.94, "flops_per_problem": 13157945671.68}, "x.parity.count24": {"acc": 0.892, "reasoning_bytes": 202.54, "flops_per_problem": 10932126842.88}, "x.parity.count28": {"acc": 0.996, "reasoning_bytes": 223.17, "flops_per_problem": 12019863306.24}, "x.parity.count32": {"acc": 1.0, "reasoning_bytes": 243.19, "flops_per_problem": 13080660295.68}, "x.parity.find24": {"acc": 0.948, "reasoning_bytes": 201.634, "flops_per_problem": 10847952027.648}, "x.parity.find28": {"acc": 0.814, "reasoning_bytes": 222.194, "flops_per_problem": 11932597075.968}, "x.parity.find32": {"acc": 0.854, "reasoning_bytes": 241.838, "flops_per_problem": 12976435445.76}, "x.parity.rev24": {"acc": 0.694, "reasoning_bytes": 195.028, "flops_per_problem": 10423721558.016}, "x.parity.rev28": {"acc": 0.4, "reasoning_bytes": 216.644, "flops_per_problem": 11555002810.368}, "x.parity.rev32": {"acc": 0.47, "reasoning_bytes": 237.564, "flops_per_problem": 12655546564.608}, "x.parity.sumd24": {"acc": 0.746, "reasoning_bytes": 199.316, "flops_per_problem": 10657255882.752}, "x.parity.sumd28": {"acc": 0.416, "reasoning_bytes": 220.64, "flops_per_problem": 11775641518.08}, "x.parity.sumd32": {"acc": 0.494, "reasoning_bytes": 240.97, "flops_per_problem": 12849864081.408}, "x.parity.inc24": {"acc": 0.776, "reasoning_bytes": 199.898, "flops_per_problem": 10638795718.656}, "x.parity.inc28": {"acc": 0.382, "reasoning_bytes": 220.224, "flops_per_problem": 11713106608.128}, "x.parity.inc32": {"acc": 0.472, "reasoning_bytes": 240.544, "flops_per_problem": 12787152519.168}, "x.sumd.cmp8": {"acc": 0.992, "reasoning_bytes": 101.09, "flops_per_problem": 5612596494.336}, "x.sumd.cmp9": {"acc": 0.98, "reasoning_bytes": 109.896, "flops_per_problem": 6045747904.512}, "x.sumd.cmp10": {"acc": 0.914, "reasoning_bytes": 116.718, "flops_per_problem": 6392163041.28}, "x.sumd.runmax8": {"acc": 0.992, "reasoning_bytes": 109.582, "flops_per_problem": 5987717627.904}, "x.sumd.runmax9": {"acc": 0.952, "reasoning_bytes": 118.35, "flops_per_problem": 6419190841.344}, "x.sumd.runmax10": {"acc": 0.866, "reasoning_bytes": 126.83, "flops_per_problem": 6837945090.048}, "x.sumd.count8": {"acc": 0.992, "reasoning_bytes": 107.618, "flops_per_problem": 5945144426.496}, "x.sumd.count9": {"acc": 0.934, "reasoning_bytes": 116.3, "flops_per_problem": 6372731289.6}, "x.sumd.count10": {"acc": 0.824, "reasoning_bytes": 124.8, "flops_per_problem": 6792545452.032}, "x.sumd.find8": {"acc": 0.982, "reasoning_bytes": 106.298, "flops_per_problem": 5842686099.456}, "x.sumd.find9": {"acc": 0.94, "reasoning_bytes": 114.888, "flops_per_problem": 6266209959.936}, "x.sumd.find10": {"acc": 0.808, "reasoning_bytes": 122.32, "flops_per_problem": 6638681309.184}, "x.sumd.rev8": {"acc": 0.948, "reasoning_bytes": 100.646, "flops_per_problem": 5504573620.224}, "x.sumd.rev9": {"acc": 0.8, "reasoning_bytes": 109.396, "flops_per_problem": 5934633615.36}, "x.sumd.rev10": {"acc": 0.616, "reasoning_bytes": 118.0, "flops_per_problem": 6358334128.128}, "x.sumd.sumd8": {"acc": 0.828, "reasoning_bytes": 105.852, "flops_per_problem": 5740051120.128}, "x.sumd.sumd9": {"acc": 0.65, "reasoning_bytes": 114.566, "flops_per_problem": 6167461330.944}, "x.sumd.sumd10": {"acc": 0.482, "reasoning_bytes": 122.43, "flops_per_problem": 6559276105.728}, "x.sumd.inc8": {"acc": 0.2, "reasoning_bytes": 108.14, "flops_per_problem": 5834030137.344}, "x.sumd.inc9": {"acc": 0.104, "reasoning_bytes": 118.76, "flops_per_problem": 6339608985.6}, "x.sumd.inc10": {"acc": 0.036, "reasoning_bytes": 128.418, "flops_per_problem": 6808974114.816}, "x.inc.cmp8": {"acc": 0.996, "reasoning_bytes": 122.15, "flops_per_problem": 6763574476.8}, "x.inc.cmp9": {"acc": 0.982, "reasoning_bytes": 133.882, "flops_per_problem": 7370021781.504}, "x.inc.cmp10": {"acc": 0.918, "reasoning_bytes": 143.632, "flops_per_problem": 7891852640.256}, "x.inc.runmax8": {"acc": 0.998, "reasoning_bytes": 159.988, "flops_per_problem": 8169638363.136}, "x.inc.runmax9": {"acc": 0.998, "reasoning_bytes": 176.748, "flops_per_problem": 8953974521.856}, "x.inc.runmax10": {"acc": 0.966, "reasoning_bytes": 184.14, "flops_per_problem": 9324591022.08}, "x.inc.count8": {"acc": 0.902, "reasoning_bytes": 151.974, "flops_per_problem": 7859878576.128}, "x.inc.count9": {"acc": 0.832, "reasoning_bytes": 167.256, "flops_per_problem": 8578941714.432}, "x.inc.count10": {"acc": 0.702, "reasoning_bytes": 172.09, "flops_per_problem": 8836589076.48}, "x.inc.find8": {"acc": 0.896, "reasoning_bytes": 138.552, "flops_per_problem": 7222958751.744}, "x.inc.find9": {"acc": 0.76, "reasoning_bytes": 152.968, "flops_per_problem": 7903776669.696}, "x.inc.find10": {"acc": 0.582, "reasoning_bytes": 156.03, "flops_per_problem": 8083167068.16}, "x.inc.rev8": {"acc": 0.852, "reasoning_bytes": 165.922, "flops_per_problem": 8607471058.944}, "x.inc.rev9": {"acc": 0.518, "reasoning_bytes": 187.56, "flops_per_problem": 9647511404.544}, "x.inc.rev10": {"acc": 0.032, "reasoning_bytes": 193.838, "flops_per_problem": 9978734444.544}, "x.inc.sumd8": {"acc": 0.114, "reasoning_bytes": 167.198, "flops_per_problem": 8443537735.68}, "x.inc.sumd9": {"acc": 0.034, "reasoning_bytes": 184.32, "flops_per_problem": 9243242643.456}, "x.inc.sumd10": {"acc": 0.04, "reasoning_bytes": 190.114, "flops_per_problem": 9542756597.76}, "x.inc.inc8": {"acc": 0.836, "reasoning_bytes": 169.658, "flops_per_problem": 8773965840.384}, "x.inc.inc9": {"acc": 0.356, "reasoning_bytes": 184.902, "flops_per_problem": 9532157460.48}, "x.inc.inc10": {"acc": 0.016, "reasoning_bytes": 191.224, "flops_per_problem": 9872654745.6}} +{"step": 8000, "train_loss": 1.0680303573608398} +{"step": 8020, "train_loss": 1.0087237358093262} +{"step": 8040, "train_loss": 1.0609989166259766} +{"step": 8060, "train_loss": 1.1116608381271362} +{"step": 8080, "train_loss": 1.1702790260314941} +{"step": 8100, "train_loss": 1.1559299230575562} +{"step": 8120, "train_loss": 1.0545060634613037} +{"step": 8140, "train_loss": 1.0787907838821411} +{"step": 8160, "train_loss": 0.9686641693115234} +{"step": 8180, "train_loss": 1.0309877395629883} +{"step": 8200, "train_loss": 1.1072170734405518} +{"step": 8220, "train_loss": 1.0204697847366333} +{"step": 8240, "train_loss": 1.0143611431121826} +{"step": 8260, "train_loss": 1.1018465757369995} +{"step": 8280, "train_loss": 1.0033124685287476} +{"step": 8300, "train_loss": 1.0356637239456177} +{"step": 8320, "train_loss": 1.0482535362243652} +{"step": 8340, "train_loss": 1.0528538227081299} +{"step": 8360, "train_loss": 1.0747517347335815} +{"step": 8380, "train_loss": 1.0135865211486816} +{"step": 8400, "train_loss": 1.094773530960083} +{"step": 8420, "train_loss": 1.0728983879089355} +{"step": 8440, "train_loss": 1.0429636240005493} +{"step": 8460, "train_loss": 0.9689235091209412} +{"step": 8480, "train_loss": 1.0425982475280762} +{"step": 8500, "train_loss": 1.019277811050415} +{"step": 8520, "train_loss": 1.1173934936523438} +{"step": 8540, "train_loss": 1.1589323282241821} +{"step": 8560, "train_loss": 1.0483007431030273} +{"step": 8580, "train_loss": 0.9975913166999817} +{"step": 8600, "train_loss": 0.9939274787902832} +{"step": 8620, "train_loss": 0.9744576215744019} +{"step": 8640, "train_loss": 1.0196377038955688} +{"step": 8660, "train_loss": 1.0692381858825684} +{"step": 8680, "train_loss": 1.0215951204299927} +{"step": 8700, "train_loss": 0.9773880243301392} +{"step": 8720, "train_loss": 0.9360859394073486} +{"step": 8740, "train_loss": 1.0465881824493408} +{"step": 8760, "train_loss": 1.0588953495025635} +{"step": 8780, "train_loss": 1.0607396364212036} +{"step": 8800, "train_loss": 1.0372719764709473} +{"step": 8820, "train_loss": 0.9769346117973328} +{"step": 8840, "train_loss": 1.1293914318084717} +{"step": 8860, "train_loss": 0.9783117771148682} +{"step": 8880, "train_loss": 1.0389430522918701} +{"step": 8900, "train_loss": 1.0288102626800537} +{"step": 8920, "train_loss": 1.0542707443237305} +{"step": 8940, "train_loss": 1.0003528594970703} +{"step": 8960, "train_loss": 1.0959742069244385} +{"step": 8980, "train_loss": 1.0504443645477295} +{"step": 9000, "train_loss": 1.05210280418396} +{"step": 9020, "train_loss": 0.9896107912063599} +{"step": 9040, "train_loss": 1.07041597366333} +{"step": 9060, "train_loss": 1.0033001899719238} +{"step": 9080, "train_loss": 0.9926108121871948} +{"step": 9100, "train_loss": 1.0077996253967285} +{"step": 9120, "train_loss": 0.9640330076217651} +{"step": 9140, "train_loss": 1.0061882734298706} +{"step": 9160, "train_loss": 0.9808368682861328} +{"step": 9180, "train_loss": 0.9942547678947449} +{"step": 9200, "train_loss": 0.9597593545913696} +{"step": 9220, "train_loss": 1.014148235321045} +{"step": 9240, "train_loss": 0.9609102010726929} +{"step": 9260, "train_loss": 0.9989566802978516} +{"step": 9280, "train_loss": 1.1036207675933838} +{"step": 9300, "train_loss": 1.1077102422714233} +{"step": 9320, "train_loss": 1.036210536956787} +{"step": 9340, "train_loss": 0.9660742282867432} +{"step": 9360, "train_loss": 0.9807233810424805} +{"step": 9380, "train_loss": 0.9633151292800903} +{"step": 9400, "train_loss": 1.0433051586151123} +{"step": 9420, "train_loss": 1.0603814125061035} +{"step": 9440, "train_loss": 0.9965790510177612} +{"step": 9460, "train_loss": 1.0648326873779297} +{"step": 9480, "train_loss": 1.0083884000778198} +{"step": 9500, "train_loss": 1.0050489902496338} +{"step": 9520, "train_loss": 0.9886621236801147} +{"step": 9540, "train_loss": 0.9966225624084473} +{"step": 9560, "train_loss": 0.9481309056282043} +{"step": 9580, "train_loss": 1.032731056213379} +{"step": 9600, "train_loss": 1.0244578123092651} +{"step": 9620, "train_loss": 0.9744634628295898} +{"step": 9640, "train_loss": 1.1282713413238525} +{"step": 9660, "train_loss": 1.0372638702392578} +{"step": 9680, "train_loss": 0.9771703481674194} +{"step": 9700, "train_loss": 1.0119472742080688} +{"step": 9720, "train_loss": 0.9744781255722046} +{"step": 9740, "train_loss": 1.0863935947418213} +{"step": 9760, "train_loss": 0.9507960081100464} +{"step": 9780, "train_loss": 0.9137444496154785} +{"step": 9800, "train_loss": 1.1014599800109863} +{"step": 9820, "train_loss": 0.961419403553009} +{"step": 9840, "train_loss": 1.0054569244384766} +{"step": 9860, "train_loss": 1.1371464729309082} +{"step": 9880, "train_loss": 1.0456453561782837} +{"step": 9900, "train_loss": 0.9489133954048157} +{"step": 9920, "train_loss": 1.0442609786987305} +{"step": 9940, "train_loss": 1.0738104581832886} +{"step": 9960, "train_loss": 1.023882269859314} +{"step": 9980, "train_loss": 0.9533258676528931} +{"step": 10000, "train_loss": 1.0205379724502563} +{"step": 10020, "train_loss": 0.9681707620620728} +{"step": 10040, "train_loss": 0.9964028596878052} +{"step": 10060, "train_loss": 0.9661391973495483} +{"step": 10080, "train_loss": 0.9718766212463379} +{"step": 10100, "train_loss": 1.0359539985656738} +{"step": 10120, "train_loss": 0.9547041654586792} +{"step": 10140, "train_loss": 0.9933766722679138} +{"step": 10160, "train_loss": 0.9878540635108948} +{"step": 10180, "train_loss": 1.049026608467102} +{"step": 10200, "train_loss": 1.0013539791107178} +{"step": 10220, "train_loss": 1.0721800327301025} +{"step": 10240, "train_loss": 1.063504695892334} +{"step": 10260, "train_loss": 1.0612187385559082} +{"step": 10280, "train_loss": 0.9837343692779541} +{"step": 10300, "train_loss": 1.0312495231628418} +{"step": 10320, "train_loss": 0.9523121118545532} +{"step": 10340, "train_loss": 1.023724913597107} +{"step": 10360, "train_loss": 1.0001225471496582} +{"step": 10380, "train_loss": 1.0490458011627197} +{"step": 10400, "train_loss": 0.9465834498405457} +{"step": 10420, "train_loss": 1.0421286821365356} +{"step": 10440, "train_loss": 1.0107520818710327} +{"step": 10460, "train_loss": 1.0560778379440308} +{"step": 10480, "train_loss": 1.057691216468811} +{"step": 10500, "train_loss": 0.9813089966773987} +{"step": 10520, "train_loss": 1.0312800407409668} +{"step": 10540, "train_loss": 0.9495437741279602} +{"step": 10560, "train_loss": 1.0135157108306885} +{"step": 10580, "train_loss": 1.0100774765014648} +{"step": 10600, "train_loss": 1.0161559581756592} +{"step": 10620, "train_loss": 1.0392937660217285} +{"step": 10640, "train_loss": 1.0320649147033691} +{"step": 10660, "train_loss": 1.0099048614501953} +{"step": 10680, "train_loss": 0.974288821220398} +{"step": 10700, "train_loss": 1.0253612995147705} +{"step": 10720, "train_loss": 0.9473544359207153} +{"step": 10740, "train_loss": 0.9964371919631958} +{"step": 10760, "train_loss": 0.986757755279541} +{"step": 10780, "train_loss": 0.9964929819107056} +{"step": 10800, "train_loss": 0.9567748308181763} +{"step": 10820, "train_loss": 1.0018608570098877} +{"step": 10840, "train_loss": 1.0052474737167358} +{"step": 10860, "train_loss": 1.0093858242034912} +{"step": 10880, "train_loss": 1.0435823202133179} +{"step": 10900, "train_loss": 1.0341168642044067} +{"step": 10920, "train_loss": 0.9806206226348877} +{"step": 10940, "train_loss": 1.0648157596588135} +{"step": 10960, "train_loss": 0.9937769770622253} +{"step": 10980, "train_loss": 0.9822419881820679} +{"step": 11000, "train_loss": 1.035556435585022} +{"step": 11020, "train_loss": 0.983606219291687} +{"step": 11040, "train_loss": 0.9490998387336731} +{"step": 11060, "train_loss": 0.9993434548377991} +{"step": 11080, "train_loss": 1.025671362876892} +{"step": 11100, "train_loss": 1.097069501876831} +{"step": 11120, "train_loss": 1.0448601245880127} +{"step": 11140, "train_loss": 0.988253116607666} +{"step": 11160, "train_loss": 1.0146586894989014} +{"step": 11180, "train_loss": 1.0515313148498535} +{"step": 11200, "train_loss": 0.9798600077629089} +{"step": 11220, "train_loss": 0.9626961946487427} +{"step": 11240, "train_loss": 1.003291130065918} +{"step": 11260, "train_loss": 1.08207368850708} +{"step": 11280, "train_loss": 0.9793714284896851} +{"step": 11300, "train_loss": 1.0279550552368164} +{"step": 11320, "train_loss": 1.016789436340332} +{"step": 11340, "train_loss": 1.0137238502502441} +{"step": 11360, "train_loss": 1.0544426441192627} +{"step": 11380, "train_loss": 1.0052298307418823} +{"step": 11400, "train_loss": 1.1014463901519775} +{"step": 11420, "train_loss": 0.971957802772522} +{"step": 11440, "train_loss": 1.0355620384216309} +{"step": 11460, "train_loss": 0.9625515937805176} +{"step": 11480, "train_loss": 0.9265741109848022} +{"step": 11500, "train_loss": 0.9705498218536377} +{"step": 11520, "train_loss": 1.0016916990280151} +{"step": 11540, "train_loss": 1.0072832107543945} +{"step": 11560, "train_loss": 0.973003625869751} +{"step": 11580, "train_loss": 1.0721018314361572} +{"step": 11600, "train_loss": 1.0262508392333984} +{"step": 11620, "train_loss": 0.9663128852844238} +{"step": 11640, "train_loss": 1.0497896671295166} +{"step": 11660, "train_loss": 1.0385868549346924} +{"step": 11680, "train_loss": 1.0093389749526978} +{"step": 11700, "train_loss": 0.9749834537506104} +{"step": 11720, "train_loss": 1.0419502258300781} +{"step": 11740, "train_loss": 1.0428497791290283} +{"step": 11760, "train_loss": 1.036449670791626} +{"step": 11780, "train_loss": 0.9743242263793945} +{"step": 11800, "train_loss": 0.9276347160339355} +{"step": 11820, "train_loss": 1.0107345581054688} +{"step": 11840, "train_loss": 0.9834208488464355} +{"step": 11860, "train_loss": 1.0370694398880005} +{"step": 11880, "train_loss": 0.9774352312088013} +{"step": 11900, "train_loss": 1.023092269897461} +{"step": 11920, "train_loss": 0.9403728246688843} +{"step": 11940, "train_loss": 0.9818663597106934} +{"step": 11960, "train_loss": 0.9295467734336853} +{"step": 11980, "train_loss": 1.0302411317825317} +{"step": 12000, "tokens": 393216000, "val_loss": 1.8186707377433777, "val_bpb": 0.6556012543670425, "lr": 6.039882929516903e-05, "elapsed": 853.6967422962189, "peak_mem_gb": 2.254330368, "add8": {"acc": 1.0, "reasoning_bytes": 79.0, "flops_per_problem": 4928775487.488}, "add9": {"acc": 0.94, "reasoning_bytes": 89.0, "flops_per_problem": 5503072075.776}, "add10": {"acc": 0.322, "reasoning_bytes": 99.0, "flops_per_problem": 6077986947.072}, "reach14": {"acc": 0.574, "reasoning_bytes": 7.932, "flops_per_problem": 4634119471.104}, "reach16": {"acc": 0.566, "reasoning_bytes": 7.992, "flops_per_problem": 5164429639.68}, "reach18": {"acc": 0.524, "reasoning_bytes": 8.056, "flops_per_problem": 5701010964.48}, "parity24": {"acc": 0.98, "reasoning_bytes": 143.012, "flops_per_problem": 7861556772.864}, "parity28": {"acc": 0.858, "reasoning_bytes": 163.94, "flops_per_problem": 8962453831.68}, "parity32": {"acc": 0.594, "reasoning_bytes": 185.04, "flops_per_problem": 10070946938.88}, "cmp8": {"acc": 0.998, "reasoning_bytes": 5.836, "flops_per_problem": 1362077466.624}, "cmp9": {"acc": 0.994, "reasoning_bytes": 5.72, "flops_per_problem": 1445015715.84}, "cmp10": {"acc": 0.998, "reasoning_bytes": 5.78, "flops_per_problem": 1535991644.16}, "runmax12": {"acc": 1.0, "reasoning_bytes": 71.0, "flops_per_problem": 4151328768.0}, "runmax14": {"acc": 1.0, "reasoning_bytes": 82.988, "flops_per_problem": 4769081819.136}, "runmax16": {"acc": 0.978, "reasoning_bytes": 89.274, "flops_per_problem": 5134840381.44}, "count16": {"acc": 0.992, "reasoning_bytes": 79.042, "flops_per_problem": 4727215226.88}, "count18": {"acc": 0.986, "reasoning_bytes": 88.902, "flops_per_problem": 5250989260.8}, "count20": {"acc": 0.934, "reasoning_bytes": 97.768, "flops_per_problem": 5730865201.152}, "brackets16": {"acc": 0.97, "reasoning_bytes": 54.008, "flops_per_problem": 3666064932.864}, "brackets18": {"acc": 0.968, "reasoning_bytes": 57.49, "flops_per_problem": 3907990241.28}, "brackets20": {"acc": 0.942, "reasoning_bytes": 64.686, "flops_per_problem": 4314113851.392}, "sumd8": {"acc": 0.996, "reasoning_bytes": 58.578, "flops_per_problem": 3381831401.472}, "sumd9": {"acc": 0.964, "reasoning_bytes": 66.432, "flops_per_problem": 3772851240.96}, "sumd10": {"acc": 0.77, "reasoning_bytes": 72.872, "flops_per_problem": 4097803124.736}, "inc8": {"acc": 1.0, "reasoning_bytes": 63.0, "flops_per_problem": 3798024192.0}, "inc9": {"acc": 0.992, "reasoning_bytes": 71.0, "flops_per_problem": 4239654912.0}, "inc10": {"acc": 0.392, "reasoning_bytes": 75.56, "flops_per_problem": 4510374543.36}, "add.cmp6": {"acc": 1.0, "reasoning_bytes": 92.622, "flops_per_problem": 5532043051.008}, "add.cmp7": {"acc": 0.998, "reasoning_bytes": 105.526, "flops_per_problem": 6231674437.632}, "add.cmp8": {"acc": 0.982, "reasoning_bytes": 118.652, "flops_per_problem": 6945349681.152}, "add.runmax6": {"acc": 1.0, "reasoning_bytes": 120.832, "flops_per_problem": 6572878331.904}, "add.runmax7": {"acc": 1.0, "reasoning_bytes": 138.848, "flops_per_problem": 7456846381.056}, "add.runmax8": {"acc": 0.974, "reasoning_bytes": 156.856, "flops_per_problem": 8340461125.632}, "add.count6": {"acc": 0.994, "reasoning_bytes": 115.648, "flops_per_problem": 6388100038.656}, "add.count7": {"acc": 0.988, "reasoning_bytes": 132.088, "flops_per_problem": 7202467086.336}, "add.count8": {"acc": 0.866, "reasoning_bytes": 148.954, "flops_per_problem": 8035647602.688}, "add.find6": {"acc": 1.0, "reasoning_bytes": 107.696, "flops_per_problem": 5992752218.112}, "add.find7": {"acc": 0.968, "reasoning_bytes": 121.2, "flops_per_problem": 6677368160.256}, "add.find8": {"acc": 0.828, "reasoning_bytes": 134.736, "flops_per_problem": 7363485646.848}, "add.rev6": {"acc": 1.0, "reasoning_bytes": 123.248, "flops_per_problem": 6794576953.344}, "add.rev7": {"acc": 1.0, "reasoning_bytes": 145.878, "flops_per_problem": 7926564814.848}, "add.rev8": {"acc": 0.87, "reasoning_bytes": 169.576, "flops_per_problem": 9106072141.824}, "add.sumd6": {"acc": 1.0, "reasoning_bytes": 126.476, "flops_per_problem": 6777794985.984}, "add.sumd7": {"acc": 0.994, "reasoning_bytes": 146.346, "flops_per_problem": 7743729696.768}, "add.sumd8": {"acc": 0.816, "reasoning_bytes": 166.384, "flops_per_problem": 8717083803.648}, "add.inc6": {"acc": 1.0, "reasoning_bytes": 131.04, "flops_per_problem": 7138695610.368}, "add.inc7": {"acc": 1.0, "reasoning_bytes": 151.06, "flops_per_problem": 8155417853.952}, "add.inc8": {"acc": 0.97, "reasoning_bytes": 171.04, "flops_per_problem": 9170285248.512}, "runmax.cmp12": {"acc": 1.0, "reasoning_bytes": 87.0, "flops_per_problem": 5122916352.0}, "runmax.cmp14": {"acc": 1.0, "reasoning_bytes": 98.916, "flops_per_problem": 5737489661.952}, "runmax.cmp16": {"acc": 0.992, "reasoning_bytes": 105.396, "flops_per_problem": 6111727534.08}, "runmax.runmax12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5211242496.0}, "runmax.runmax14": {"acc": 0.996, "reasoning_bytes": 99.22, "flops_per_problem": 5795078307.84}, "runmax.runmax16": {"acc": 0.97, "reasoning_bytes": 105.928, "flops_per_problem": 6179473686.528}, "runmax.count12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5255405568.0}, "runmax.count14": {"acc": 1.0, "reasoning_bytes": 99.976, "flops_per_problem": 5872628662.272}, "runmax.count16": {"acc": 0.992, "reasoning_bytes": 106.13, "flops_per_problem": 6232027742.208}, "runmax.find12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5211242496.0}, "runmax.find14": {"acc": 1.0, "reasoning_bytes": 99.928, "flops_per_problem": 5826345762.816}, "runmax.find16": {"acc": 0.994, "reasoning_bytes": 106.18, "flops_per_problem": 6189012910.08}, "runmax.rev12": {"acc": 1.0, "reasoning_bytes": 83.0, "flops_per_problem": 4857937920.0}, "runmax.rev14": {"acc": 1.0, "reasoning_bytes": 94.868, "flops_per_problem": 5470391402.496}, "runmax.rev16": {"acc": 0.97, "reasoning_bytes": 101.052, "flops_per_problem": 5831733657.6}, "runmax.sumd12": {"acc": 1.0, "reasoning_bytes": 86.0, "flops_per_problem": 5034590208.0}, "runmax.sumd14": {"acc": 0.996, "reasoning_bytes": 96.956, "flops_per_problem": 5606766968.832}, "runmax.sumd16": {"acc": 0.966, "reasoning_bytes": 103.962, "flops_per_problem": 6004322942.976}, "runmax.inc12": {"acc": 1.0, "reasoning_bytes": 87.0, "flops_per_problem": 5065769336.832}, "runmax.inc14": {"acc": 1.0, "reasoning_bytes": 98.916, "flops_per_problem": 5683699040.256}, "runmax.inc16": {"acc": 0.978, "reasoning_bytes": 105.296, "flops_per_problem": 6054933823.488}, "count.cmp16": {"acc": 0.994, "reasoning_bytes": 95.01, "flops_per_problem": 5697477918.72}, "count.cmp18": {"acc": 0.99, "reasoning_bytes": 104.928, "flops_per_problem": 6224608346.112}, "count.cmp20": {"acc": 0.982, "reasoning_bytes": 113.472, "flops_per_problem": 6689645494.272}, "count.runmax16": {"acc": 0.998, "reasoning_bytes": 96.01, "flops_per_problem": 5785804062.72}, "count.runmax18": {"acc": 0.98, "reasoning_bytes": 105.756, "flops_per_problem": 6304543506.432}, "count.runmax20": {"acc": 0.882, "reasoning_bytes": 113.646, "flops_per_problem": 6741316288.512}, "count.count16": {"acc": 1.0, "reasoning_bytes": 96.0, "flops_per_problem": 5829525504.0}, "count.count18": {"acc": 1.0, "reasoning_bytes": 105.946, "flops_per_problem": 6357097562.112}, "count.count20": {"acc": 0.998, "reasoning_bytes": 114.552, "flops_per_problem": 6825491103.744}, "count.find16": {"acc": 0.998, "reasoning_bytes": 95.99, "flops_per_problem": 5784920801.28}, "count.find18": {"acc": 0.996, "reasoning_bytes": 105.81, "flops_per_problem": 6306928312.32}, "count.find20": {"acc": 0.984, "reasoning_bytes": 114.29, "flops_per_problem": 6769757306.88}, "count.rev16": {"acc": 0.996, "reasoning_bytes": 91.016, "flops_per_problem": 5432676139.008}, "count.rev18": {"acc": 0.988, "reasoning_bytes": 100.896, "flops_per_problem": 5957421760.512}, "count.rev20": {"acc": 0.908, "reasoning_bytes": 109.306, "flops_per_problem": 6417159340.032}, "count.sumd16": {"acc": 0.996, "reasoning_bytes": 94.016, "flops_per_problem": 5609328427.008}, "count.sumd18": {"acc": 0.976, "reasoning_bytes": 103.956, "flops_per_problem": 6136635506.688}, "count.sumd20": {"acc": 0.9, "reasoning_bytes": 112.312, "flops_per_problem": 6594076606.464}, "count.inc16": {"acc": 0.996, "reasoning_bytes": 95.012, "flops_per_problem": 5609505079.296}, "count.inc18": {"acc": 0.982, "reasoning_bytes": 104.924, "flops_per_problem": 6135398940.672}, "count.inc20": {"acc": 0.91, "reasoning_bytes": 113.456, "flops_per_problem": 6600436088.832}, "parity.cmp16": {"acc": 1.0, "reasoning_bytes": 111.0, "flops_per_problem": 6359482368.0}, "parity.cmp20": {"acc": 1.0, "reasoning_bytes": 134.868, "flops_per_problem": 7590218858.496}, "parity.cmp24": {"acc": 0.966, "reasoning_bytes": 154.56, "flops_per_problem": 8636530360.32}, "parity.runmax16": {"acc": 1.0, "reasoning_bytes": 112.0, "flops_per_problem": 6447808512.0}, "parity.runmax20": {"acc": 0.996, "reasoning_bytes": 135.832, "flops_per_problem": 7676955131.904}, "parity.runmax24": {"acc": 0.576, "reasoning_bytes": 154.348, "flops_per_problem": 8671330861.056}, "parity.count16": {"acc": 1.0, "reasoning_bytes": 112.0, "flops_per_problem": 6491971584.0}, "parity.count20": {"acc": 1.0, "reasoning_bytes": 135.964, "flops_per_problem": 7726947729.408}, "parity.count24": {"acc": 0.992, "reasoning_bytes": 154.576, "flops_per_problem": 8725563113.472}, "parity.find16": {"acc": 1.0, "reasoning_bytes": 112.0, "flops_per_problem": 6447808512.0}, "parity.find20": {"acc": 0.998, "reasoning_bytes": 135.94, "flops_per_problem": 7681724743.68}, "parity.find24": {"acc": 0.916, "reasoning_bytes": 155.032, "flops_per_problem": 8701538402.304}, "parity.rev16": {"acc": 1.0, "reasoning_bytes": 107.0, "flops_per_problem": 6094503936.0}, "parity.rev20": {"acc": 0.998, "reasoning_bytes": 130.964, "flops_per_problem": 7329480081.408}, "parity.rev24": {"acc": 0.666, "reasoning_bytes": 150.212, "flops_per_problem": 8356183179.264}, "parity.sumd16": {"acc": 1.0, "reasoning_bytes": 110.0, "flops_per_problem": 6271156224.0}, "parity.sumd20": {"acc": 0.988, "reasoning_bytes": 133.796, "flops_per_problem": 7498712973.312}, "parity.sumd24": {"acc": 0.614, "reasoning_bytes": 152.864, "flops_per_problem": 8517466718.208}, "parity.inc16": {"acc": 1.0, "reasoning_bytes": 111.0, "flops_per_problem": 6271156224.0}, "parity.inc20": {"acc": 0.996, "reasoning_bytes": 134.892, "flops_per_problem": 7502952628.224}, "parity.inc24": {"acc": 0.738, "reasoning_bytes": 154.788, "flops_per_problem": 8558273396.736}, "sumd.cmp6": {"acc": 1.0, "reasoning_bytes": 62.088, "flops_per_problem": 3713407746.048}, "sumd.cmp7": {"acc": 1.0, "reasoning_bytes": 70.084, "flops_per_problem": 4110875394.048}, "sumd.cmp8": {"acc": 0.996, "reasoning_bytes": 77.888, "flops_per_problem": 4499598753.792}, "sumd.runmax6": {"acc": 1.0, "reasoning_bytes": 67.542, "flops_per_problem": 3954449793.024}, "sumd.runmax7": {"acc": 1.0, "reasoning_bytes": 75.48, "flops_per_problem": 4349179330.56}, "sumd.runmax8": {"acc": 0.984, "reasoning_bytes": 83.422, "flops_per_problem": 4744085520.384}, "sumd.count6": {"acc": 1.0, "reasoning_bytes": 66.428, "flops_per_problem": 3949415202.816}, "sumd.count7": {"acc": 1.0, "reasoning_bytes": 74.46, "flops_per_problem": 4348296069.12}, "sumd.count8": {"acc": 0.996, "reasoning_bytes": 82.588, "flops_per_problem": 4751416590.336}, "sumd.find6": {"acc": 1.0, "reasoning_bytes": 66.24, "flops_per_problem": 3896949473.28}, "sumd.find7": {"acc": 1.0, "reasoning_bytes": 74.428, "flops_per_problem": 4302719778.816}, "sumd.find8": {"acc": 0.996, "reasoning_bytes": 82.308, "flops_per_problem": 4694887858.176}, "sumd.rev6": {"acc": 1.0, "reasoning_bytes": 61.544, "flops_per_problem": 3601145217.024}, "sumd.rev7": {"acc": 1.0, "reasoning_bytes": 69.48, "flops_per_problem": 3995874754.56}, "sumd.rev8": {"acc": 0.994, "reasoning_bytes": 77.52, "flops_per_problem": 4395020599.296}, "sumd.sumd6": {"acc": 1.0, "reasoning_bytes": 65.744, "flops_per_problem": 3795639386.112}, "sumd.sumd7": {"acc": 1.0, "reasoning_bytes": 73.738, "flops_per_problem": 4195315187.712}, "sumd.sumd8": {"acc": 0.99, "reasoning_bytes": 81.772, "flops_per_problem": 4595520946.176}, "sumd.inc6": {"acc": 1.0, "reasoning_bytes": 68.538, "flops_per_problem": 3910021742.592}, "sumd.inc7": {"acc": 0.984, "reasoning_bytes": 76.48, "flops_per_problem": 4305016258.56}, "sumd.inc8": {"acc": 0.898, "reasoning_bytes": 84.482, "flops_per_problem": 4702572232.704}, "inc.cmp6": {"acc": 0.996, "reasoning_bytes": 78.624, "flops_per_problem": 4576354172.928}, "inc.cmp7": {"acc": 0.996, "reasoning_bytes": 89.792, "flops_per_problem": 5157893505.024}, "inc.cmp8": {"acc": 0.99, "reasoning_bytes": 100.882, "flops_per_problem": 5735988117.504}, "inc.runmax6": {"acc": 1.0, "reasoning_bytes": 104.0, "flops_per_problem": 5520384000.0}, "inc.runmax7": {"acc": 1.0, "reasoning_bytes": 120.0, "flops_per_problem": 6271156224.0}, "inc.runmax8": {"acc": 0.998, "reasoning_bytes": 135.816, "flops_per_problem": 7013802442.752}, "inc.count6": {"acc": 0.988, "reasoning_bytes": 99.0, "flops_per_problem": 5343731712.0}, "inc.count7": {"acc": 0.986, "reasoning_bytes": 113.996, "flops_per_problem": 6050164211.712}, "inc.count8": {"acc": 0.942, "reasoning_bytes": 128.968, "flops_per_problem": 6755536797.696}, "inc.find6": {"acc": 1.0, "reasoning_bytes": 91.88, "flops_per_problem": 4985127567.36}, "inc.find7": {"acc": 1.0, "reasoning_bytes": 102.516, "flops_per_problem": 5499009073.152}, "inc.find8": {"acc": 0.898, "reasoning_bytes": 116.846, "flops_per_problem": 6176028966.912}, "inc.rev6": {"acc": 1.0, "reasoning_bytes": 104.0, "flops_per_problem": 5608710144.0}, "inc.rev7": {"acc": 1.0, "reasoning_bytes": 124.0, "flops_per_problem": 6580297728.0}, "inc.rev8": {"acc": 0.972, "reasoning_bytes": 145.024, "flops_per_problem": 7596666667.008}, "inc.sumd6": {"acc": 1.0, "reasoning_bytes": 109.556, "flops_per_problem": 5721502629.888}, "inc.sumd7": {"acc": 0.994, "reasoning_bytes": 127.48, "flops_per_problem": 6557244604.416}, "inc.sumd8": {"acc": 0.85, "reasoning_bytes": 145.392, "flops_per_problem": 7392368295.936}, "inc.inc6": {"acc": 1.0, "reasoning_bytes": 113.0, "flops_per_problem": 6006177792.0}, "inc.inc7": {"acc": 1.0, "reasoning_bytes": 131.0, "flops_per_problem": 6889439232.0}, "inc.inc8": {"acc": 0.98, "reasoning_bytes": 148.95, "flops_per_problem": 7770492518.4}, "x.add.cmp8": {"acc": 0.79, "reasoning_bytes": 149.194, "flops_per_problem": 8383475957.76}, "x.add.cmp9": {"acc": 0.692, "reasoning_bytes": 164.062, "flops_per_problem": 9179471167.488}, "x.add.cmp10": {"acc": 0.604, "reasoning_bytes": 174.738, "flops_per_problem": 9903568896.0}, "x.add.runmax8": {"acc": 1.0, "reasoning_bytes": 189.832, "flops_per_problem": 9885108731.904}, "x.add.runmax9": {"acc": 0.998, "reasoning_bytes": 209.864, "flops_per_problem": 10858109534.208}, "x.add.runmax10": {"acc": 0.906, "reasoning_bytes": 229.76, "flops_per_problem": 11825104158.72}, "x.add.count8": {"acc": 0.978, "reasoning_bytes": 180.934, "flops_per_problem": 9536308789.248}, "x.add.count9": {"acc": 0.966, "reasoning_bytes": 200.324, "flops_per_problem": 10480956899.328}, "x.add.count10": {"acc": 0.802, "reasoning_bytes": 219.022, "flops_per_problem": 11395044163.584}, "x.add.find8": {"acc": 0.914, "reasoning_bytes": 165.528, "flops_per_problem": 8811769430.016}, "x.add.find9": {"acc": 0.86, "reasoning_bytes": 182.408, "flops_per_problem": 9645568229.376}, "x.add.find10": {"acc": 0.768, "reasoning_bytes": 198.724, "flops_per_problem": 10454459056.128}, "x.add.rev8": {"acc": 1.0, "reasoning_bytes": 199.456, "flops_per_problem": 10513460920.32}, "x.add.rev9": {"acc": 0.958, "reasoning_bytes": 226.09, "flops_per_problem": 11822366048.256}, "x.add.rev10": {"acc": 0.222, "reasoning_bytes": 253.17, "flops_per_problem": 13143990140.928}, "x.add.sumd8": {"acc": 0.99, "reasoning_bytes": 197.38, "flops_per_problem": 10174288527.36}, "x.add.sumd9": {"acc": 0.916, "reasoning_bytes": 219.32, "flops_per_problem": 11231552471.04}, "x.add.sumd10": {"acc": 0.54, "reasoning_bytes": 241.078, "flops_per_problem": 12280778735.616}, "x.add.inc8": {"acc": 1.0, "reasoning_bytes": 201.04, "flops_per_problem": 10583415226.368}, "x.add.inc9": {"acc": 0.992, "reasoning_bytes": 223.08, "flops_per_problem": 11689435201.536}, "x.add.inc10": {"acc": 0.582, "reasoning_bytes": 242.984, "flops_per_problem": 12694498394.112}, "x.runmax.cmp12": {"acc": 1.0, "reasoning_bytes": 118.0, "flops_per_problem": 6580297728.0}, "x.runmax.cmp14": {"acc": 1.0, "reasoning_bytes": 131.982, "flops_per_problem": 7286111944.704}, "x.runmax.cmp16": {"acc": 1.0, "reasoning_bytes": 143.804, "flops_per_problem": 7896533925.888}, "x.runmax.runmax12": {"acc": 1.0, "reasoning_bytes": 122.0, "flops_per_problem": 6801113088.0}, "x.runmax.runmax14": {"acc": 1.0, "reasoning_bytes": 135.988, "flops_per_problem": 7507192283.136}, "x.runmax.runmax16": {"acc": 0.99, "reasoning_bytes": 148.354, "flops_per_problem": 8141638975.488}, "x.runmax.count12": {"acc": 1.0, "reasoning_bytes": 121.0, "flops_per_problem": 6801113088.0}, "x.runmax.count14": {"acc": 1.0, "reasoning_bytes": 134.954, "flops_per_problem": 7505690738.688}, "x.runmax.count16": {"acc": 1.0, "reasoning_bytes": 146.946, "flops_per_problem": 8123620442.112}, "x.runmax.find12": {"acc": 1.0, "reasoning_bytes": 120.0, "flops_per_problem": 6712786944.0}, "x.runmax.find14": {"acc": 1.0, "reasoning_bytes": 134.006, "flops_per_problem": 7419661074.432}, "x.runmax.find16": {"acc": 0.998, "reasoning_bytes": 146.574, "flops_per_problem": 8063911968.768}, "x.runmax.rev12": {"acc": 1.0, "reasoning_bytes": 114.0, "flops_per_problem": 6315319296.0}, "x.runmax.rev14": {"acc": 1.0, "reasoning_bytes": 128.004, "flops_per_problem": 7022105100.288}, "x.runmax.rev16": {"acc": 0.996, "reasoning_bytes": 141.024, "flops_per_problem": 7685434441.728}, "x.runmax.sumd12": {"acc": 1.0, "reasoning_bytes": 118.0, "flops_per_problem": 6536134656.0}, "x.runmax.sumd14": {"acc": 1.0, "reasoning_bytes": 131.94, "flops_per_problem": 7240094023.68}, "x.runmax.sumd16": {"acc": 0.992, "reasoning_bytes": 143.63, "flops_per_problem": 7844686479.36}, "x.runmax.inc12": {"acc": 1.0, "reasoning_bytes": 118.0, "flops_per_problem": 6523150712.832}, "x.runmax.inc14": {"acc": 1.0, "reasoning_bytes": 131.954, "flops_per_problem": 7231084756.992}, "x.runmax.inc16": {"acc": 0.988, "reasoning_bytes": 143.414, "flops_per_problem": 7827639533.568}, "x.count.cmp16": {"acc": 0.996, "reasoning_bytes": 129.998, "flops_per_problem": 7330981625.856}, "x.count.cmp18": {"acc": 0.992, "reasoning_bytes": 141.948, "flops_per_problem": 7947144806.4}, "x.count.cmp20": {"acc": 0.96, "reasoning_bytes": 152.946, "flops_per_problem": 8521264742.4}, "x.count.runmax16": {"acc": 0.992, "reasoning_bytes": 133.986, "flops_per_problem": 7551267028.992}, "x.count.runmax18": {"acc": 0.982, "reasoning_bytes": 145.844, "flops_per_problem": 8163278880.768}, "x.count.runmax20": {"acc": 0.896, "reasoning_bytes": 156.304, "flops_per_problem": 8713550757.888}, "x.count.count16": {"acc": 1.0, "reasoning_bytes": 132.954, "flops_per_problem": 7549853810.688}, "x.count.count18": {"acc": 1.0, "reasoning_bytes": 144.874, "flops_per_problem": 8164603772.928}, "x.count.count20": {"acc": 0.996, "reasoning_bytes": 155.398, "flops_per_problem": 8717702086.656}, "x.count.find16": {"acc": 0.992, "reasoning_bytes": 131.98, "flops_per_problem": 7462675906.56}, "x.count.find18": {"acc": 0.978, "reasoning_bytes": 143.802, "flops_per_problem": 8073539518.464}, "x.count.find20": {"acc": 0.966, "reasoning_bytes": 154.388, "flops_per_problem": 8629905899.52}, "x.count.rev16": {"acc": 0.994, "reasoning_bytes": 125.99, "flops_per_problem": 7065649889.28}, "x.count.rev18": {"acc": 0.988, "reasoning_bytes": 137.934, "flops_per_problem": 7681459765.248}, "x.count.rev20": {"acc": 0.924, "reasoning_bytes": 149.136, "flops_per_problem": 8259289399.296}, "x.count.sumd16": {"acc": 0.992, "reasoning_bytes": 129.988, "flops_per_problem": 7286376923.136}, "x.count.sumd18": {"acc": 0.986, "reasoning_bytes": 141.918, "flops_per_problem": 7901568516.096}, "x.count.sumd20": {"acc": 0.924, "reasoning_bytes": 152.604, "flops_per_problem": 8462174552.064}, "x.count.inc16": {"acc": 0.996, "reasoning_bytes": 130.0, "flops_per_problem": 7242920460.288}, "x.count.inc18": {"acc": 0.99, "reasoning_bytes": 141.918, "flops_per_problem": 7857405444.096}, "x.count.inc20": {"acc": 0.938, "reasoning_bytes": 152.876, "flops_per_problem": 8429935509.504}, "x.parity.cmp24": {"acc": 1.0, "reasoning_bytes": 202.012, "flops_per_problem": 10820482596.864}, "x.parity.cmp28": {"acc": 0.988, "reasoning_bytes": 230.032, "flops_per_problem": 12234584162.304}, "x.parity.cmp32": {"acc": 0.926, "reasoning_bytes": 253.592, "flops_per_problem": 13451630100.48}, "x.parity.runmax24": {"acc": 1.0, "reasoning_bytes": 206.0, "flops_per_problem": 11040768000.0}, "x.parity.runmax28": {"acc": 0.92, "reasoning_bytes": 233.518, "flops_per_problem": 12432699703.296}, "x.parity.runmax32": {"acc": 0.436, "reasoning_bytes": 255.826, "flops_per_problem": 13594541801.472}, "x.parity.count24": {"acc": 1.0, "reasoning_bytes": 205.0, "flops_per_problem": 11040768000.0}, "x.parity.count28": {"acc": 0.998, "reasoning_bytes": 232.676, "flops_per_problem": 12439677468.672}, "x.parity.count32": {"acc": 0.952, "reasoning_bytes": 259.638, "flops_per_problem": 13805199654.912}, "x.parity.find24": {"acc": 1.0, "reasoning_bytes": 204.0, "flops_per_problem": 10952441856.0}, "x.parity.find28": {"acc": 0.988, "reasoning_bytes": 231.922, "flops_per_problem": 12362215440.384}, "x.parity.find32": {"acc": 0.874, "reasoning_bytes": 253.61, "flops_per_problem": 13496676433.92}, "x.parity.rev24": {"acc": 0.752, "reasoning_bytes": 198.0, "flops_per_problem": 10554974208.0}, "x.parity.rev28": {"acc": 0.716, "reasoning_bytes": 225.916, "flops_per_problem": 11964482813.952}, "x.parity.rev32": {"acc": 0.504, "reasoning_bytes": 252.176, "flops_per_problem": 13300945698.816}, "x.parity.sumd24": {"acc": 1.0, "reasoning_bytes": 202.0, "flops_per_problem": 10775789568.0}, "x.parity.sumd28": {"acc": 0.94, "reasoning_bytes": 229.808, "flops_per_problem": 12180528562.176}, "x.parity.sumd32": {"acc": 0.452, "reasoning_bytes": 253.294, "flops_per_problem": 13397132869.632}, "x.parity.inc24": {"acc": 1.0, "reasoning_bytes": 202.0, "flops_per_problem": 10731626496.0}, "x.parity.inc28": {"acc": 0.958, "reasoning_bytes": 229.924, "flops_per_problem": 12141488406.528}, "x.parity.inc32": {"acc": 0.492, "reasoning_bytes": 253.178, "flops_per_problem": 13345020444.672}, "x.sumd.cmp8": {"acc": 1.0, "reasoning_bytes": 101.096, "flops_per_problem": 5612861472.768}, "x.sumd.cmp9": {"acc": 1.0, "reasoning_bytes": 110.252, "flops_per_problem": 6061469958.144}, "x.sumd.cmp10": {"acc": 0.992, "reasoning_bytes": 119.15, "flops_per_problem": 6498596044.8}, "x.sumd.runmax8": {"acc": 1.0, "reasoning_bytes": 109.55, "flops_per_problem": 5986304409.6}, "x.sumd.runmax9": {"acc": 0.986, "reasoning_bytes": 118.374, "flops_per_problem": 6420162428.928}, "x.sumd.runmax10": {"acc": 0.972, "reasoning_bytes": 127.44, "flops_per_problem": 6864707911.68}, "x.sumd.count8": {"acc": 1.0, "reasoning_bytes": 107.588, "flops_per_problem": 5943819534.336}, "x.sumd.count9": {"acc": 0.984, "reasoning_bytes": 116.244, "flops_per_problem": 6370258157.568}, "x.sumd.count10": {"acc": 0.954, "reasoning_bytes": 125.498, "flops_per_problem": 6823106297.856}, "x.sumd.find8": {"acc": 1.0, "reasoning_bytes": 106.33, "flops_per_problem": 5844099317.76}, "x.sumd.find9": {"acc": 1.0, "reasoning_bytes": 115.448, "flops_per_problem": 6290941280.256}, "x.sumd.find10": {"acc": 0.976, "reasoning_bytes": 124.308, "flops_per_problem": 6726389170.176}, "x.sumd.rev8": {"acc": 1.0, "reasoning_bytes": 100.552, "flops_per_problem": 5500510617.6}, "x.sumd.rev9": {"acc": 0.984, "reasoning_bytes": 109.43, "flops_per_problem": 5936841768.96}, "x.sumd.rev10": {"acc": 0.942, "reasoning_bytes": 118.53, "flops_per_problem": 6383065448.448}, "x.sumd.sumd8": {"acc": 1.0, "reasoning_bytes": 105.838, "flops_per_problem": 5746763907.072}, "x.sumd.sumd9": {"acc": 0.974, "reasoning_bytes": 114.798, "flops_per_problem": 6190161149.952}, "x.sumd.sumd10": {"acc": 0.89, "reasoning_bytes": 123.88, "flops_per_problem": 6636384829.44}, "x.sumd.inc8": {"acc": 0.362, "reasoning_bytes": 107.542, "flops_per_problem": 5810800361.472}, "x.sumd.inc9": {"acc": 0.304, "reasoning_bytes": 116.43, "flops_per_problem": 6248544731.136}, "x.sumd.inc10": {"acc": 0.28, "reasoning_bytes": 125.46, "flops_per_problem": 6693443518.464}, "x.inc.cmp8": {"acc": 1.0, "reasoning_bytes": 121.816, "flops_per_problem": 6748824010.752}, "x.inc.cmp9": {"acc": 0.992, "reasoning_bytes": 133.71, "flops_per_problem": 7362425733.12}, "x.inc.cmp10": {"acc": 0.954, "reasoning_bytes": 145.124, "flops_per_problem": 7983623503.872}, "x.inc.runmax8": {"acc": 1.0, "reasoning_bytes": 160.0, "flops_per_problem": 8170168320.0}, "x.inc.runmax9": {"acc": 1.0, "reasoning_bytes": 176.964, "flops_per_problem": 8963513745.408}, "x.inc.runmax10": {"acc": 1.0, "reasoning_bytes": 192.784, "flops_per_problem": 9706336616.448}, "x.inc.count8": {"acc": 0.99, "reasoning_bytes": 152.0, "flops_per_problem": 7861026816.0}, "x.inc.count9": {"acc": 0.982, "reasoning_bytes": 167.772, "flops_per_problem": 8601729859.584}, "x.inc.count10": {"acc": 0.91, "reasoning_bytes": 182.14, "flops_per_problem": 9280427950.08}, "x.inc.find8": {"acc": 0.958, "reasoning_bytes": 138.536, "flops_per_problem": 7222252142.592}, "x.inc.find9": {"acc": 0.846, "reasoning_bytes": 153.556, "flops_per_problem": 7929744556.032}, "x.inc.find10": {"acc": 0.798, "reasoning_bytes": 164.032, "flops_per_problem": 8436559970.304}, "x.inc.rev8": {"acc": 1.0, "reasoning_bytes": 166.0, "flops_per_problem": 8611799040.0}, "x.inc.rev9": {"acc": 0.99, "reasoning_bytes": 188.996, "flops_per_problem": 9715699187.712}, "x.inc.rev10": {"acc": 0.634, "reasoning_bytes": 211.528, "flops_per_problem": 10794691362.816}, "x.inc.sumd8": {"acc": 0.982, "reasoning_bytes": 167.582, "flops_per_problem": 8460761333.76}, "x.inc.sumd9": {"acc": 0.694, "reasoning_bytes": 186.154, "flops_per_problem": 9325385957.376}, "x.inc.sumd10": {"acc": 0.148, "reasoning_bytes": 202.676, "flops_per_problem": 10099299631.104}, "x.inc.inc8": {"acc": 1.0, "reasoning_bytes": 170.0, "flops_per_problem": 8788451328.0}, "x.inc.inc9": {"acc": 0.98, "reasoning_bytes": 188.884, "flops_per_problem": 9710576271.36}, "x.inc.inc10": {"acc": 0.676, "reasoning_bytes": 204.936, "flops_per_problem": 10499770368.0}} +{"step": 12000, "train_loss": 1.0205798149108887} +{"step": 12020, "train_loss": 1.019404649734497} +{"step": 12040, "train_loss": 0.9529318809509277} +{"step": 12060, "train_loss": 1.035186767578125} +{"step": 12080, "train_loss": 1.0506854057312012} +{"step": 12100, "train_loss": 1.0768901109695435} +{"step": 12120, "train_loss": 1.0040546655654907} +{"step": 12140, "train_loss": 0.9992959499359131} +{"step": 12160, "train_loss": 1.0332376956939697} +{"step": 12180, "train_loss": 0.9080939888954163} +{"step": 12200, "train_loss": 1.0060795545578003} +{"step": 12207, "tokens": 399998976, "val_loss": 1.806345921754837, "val_bpb": 0.651158358435283, "lr": 5.9999999999999995e-05, "elapsed": 1047.3178360462189, "peak_mem_gb": 2.254330368, "add8": {"acc": 1.0, "reasoning_bytes": 79.0, "flops_per_problem": 4928775487.488}, "add9": {"acc": 0.952, "reasoning_bytes": 89.0, "flops_per_problem": 5503160401.92}, "add10": {"acc": 0.382, "reasoning_bytes": 99.0, "flops_per_problem": 6077810294.784}, "reach14": {"acc": 0.586, "reasoning_bytes": 7.6, "flops_per_problem": 4619457331.2}, "reach16": {"acc": 0.566, "reasoning_bytes": 7.592, "flops_per_problem": 5149237542.912}, "reach18": {"acc": 0.482, "reasoning_bytes": 7.558, "flops_per_problem": 5677692862.464}, "parity24": {"acc": 0.852, "reasoning_bytes": 142.112, "flops_per_problem": 7821810008.064}, "parity28": {"acc": 0.454, "reasoning_bytes": 161.044, "flops_per_problem": 8834557575.168}, "parity32": {"acc": 0.434, "reasoning_bytes": 181.224, "flops_per_problem": 9902420656.128}, "cmp8": {"acc": 0.998, "reasoning_bytes": 5.836, "flops_per_problem": 1362077466.624}, "cmp9": {"acc": 1.0, "reasoning_bytes": 5.708, "flops_per_problem": 1444485758.976}, "cmp10": {"acc": 1.0, "reasoning_bytes": 5.756, "flops_per_problem": 1534931730.432}, "runmax12": {"acc": 1.0, "reasoning_bytes": 71.0, "flops_per_problem": 4151328768.0}, "runmax14": {"acc": 1.0, "reasoning_bytes": 83.0, "flops_per_problem": 4769611776.0}, "runmax16": {"acc": 0.992, "reasoning_bytes": 92.372, "flops_per_problem": 5271834230.784}, "count16": {"acc": 0.996, "reasoning_bytes": 79.006, "flops_per_problem": 4725625356.288}, "count18": {"acc": 0.99, "reasoning_bytes": 88.74, "flops_per_problem": 5243923169.28}, "count20": {"acc": 0.91, "reasoning_bytes": 96.806, "flops_per_problem": 5688468652.032}, "brackets16": {"acc": 0.962, "reasoning_bytes": 52.004, "flops_per_problem": 3577385484.288}, "brackets18": {"acc": 0.968, "reasoning_bytes": 53.426, "flops_per_problem": 3728511516.672}, "brackets20": {"acc": 0.976, "reasoning_bytes": 59.49, "flops_per_problem": 4084642529.28}, "sumd8": {"acc": 0.994, "reasoning_bytes": 58.566, "flops_per_problem": 3381301444.608}, "sumd9": {"acc": 0.964, "reasoning_bytes": 66.426, "flops_per_problem": 3772674588.672}, "sumd10": {"acc": 0.876, "reasoning_bytes": 74.364, "flops_per_problem": 4167315800.064}, "inc8": {"acc": 1.0, "reasoning_bytes": 63.0, "flops_per_problem": 3798024192.0}, "inc9": {"acc": 0.988, "reasoning_bytes": 71.0, "flops_per_problem": 4239654912.0}, "inc10": {"acc": 0.452, "reasoning_bytes": 76.248, "flops_per_problem": 4544556761.088}, "add.cmp6": {"acc": 0.998, "reasoning_bytes": 92.622, "flops_per_problem": 5532043051.008}, "add.cmp7": {"acc": 1.0, "reasoning_bytes": 105.548, "flops_per_problem": 6232646025.216}, "add.cmp8": {"acc": 0.98, "reasoning_bytes": 118.842, "flops_per_problem": 6953740664.832}, "add.runmax6": {"acc": 1.0, "reasoning_bytes": 120.832, "flops_per_problem": 6572878331.904}, "add.runmax7": {"acc": 1.0, "reasoning_bytes": 138.848, "flops_per_problem": 7456846381.056}, "add.runmax8": {"acc": 0.976, "reasoning_bytes": 156.832, "flops_per_problem": 8339401211.904}, "add.count6": {"acc": 0.994, "reasoning_bytes": 115.648, "flops_per_problem": 6388100038.656}, "add.count7": {"acc": 0.99, "reasoning_bytes": 132.088, "flops_per_problem": 7202467086.336}, "add.count8": {"acc": 0.746, "reasoning_bytes": 148.904, "flops_per_problem": 8033439449.088}, "add.find6": {"acc": 0.998, "reasoning_bytes": 107.696, "flops_per_problem": 5992752218.112}, "add.find7": {"acc": 0.972, "reasoning_bytes": 121.2, "flops_per_problem": 6677456486.4}, "add.find8": {"acc": 0.818, "reasoning_bytes": 135.224, "flops_per_problem": 7385125552.128}, "add.rev6": {"acc": 1.0, "reasoning_bytes": 123.248, "flops_per_problem": 6794576953.344}, "add.rev7": {"acc": 1.0, "reasoning_bytes": 145.878, "flops_per_problem": 7926564814.848}, "add.rev8": {"acc": 0.852, "reasoning_bytes": 169.456, "flops_per_problem": 9100242616.32}, "add.sumd6": {"acc": 1.0, "reasoning_bytes": 126.476, "flops_per_problem": 6777794985.984}, "add.sumd7": {"acc": 0.99, "reasoning_bytes": 146.33, "flops_per_problem": 7743023087.616}, "add.sumd8": {"acc": 0.806, "reasoning_bytes": 166.244, "flops_per_problem": 8710900973.568}, "add.inc6": {"acc": 1.0, "reasoning_bytes": 131.04, "flops_per_problem": 7138695610.368}, "add.inc7": {"acc": 1.0, "reasoning_bytes": 151.06, "flops_per_problem": 8155417853.952}, "add.inc8": {"acc": 0.93, "reasoning_bytes": 171.04, "flops_per_problem": 9171080183.808}, "runmax.cmp12": {"acc": 1.0, "reasoning_bytes": 87.0, "flops_per_problem": 5122916352.0}, "runmax.cmp14": {"acc": 1.0, "reasoning_bytes": 98.968, "flops_per_problem": 5739786141.696}, "runmax.cmp16": {"acc": 0.994, "reasoning_bytes": 107.1, "flops_per_problem": 6187246387.2}, "runmax.runmax12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5211242496.0}, "runmax.runmax14": {"acc": 0.998, "reasoning_bytes": 99.916, "flops_per_problem": 5825815805.952}, "runmax.runmax16": {"acc": 0.976, "reasoning_bytes": 106.3, "flops_per_problem": 6196079001.6}, "runmax.count12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5255405568.0}, "runmax.count14": {"acc": 1.0, "reasoning_bytes": 99.976, "flops_per_problem": 5872628662.272}, "runmax.count16": {"acc": 0.992, "reasoning_bytes": 107.23, "flops_per_problem": 6281313730.56}, "runmax.find12": {"acc": 1.0, "reasoning_bytes": 88.0, "flops_per_problem": 5211242496.0}, "runmax.find14": {"acc": 1.0, "reasoning_bytes": 99.956, "flops_per_problem": 5827582328.832}, "runmax.find16": {"acc": 0.994, "reasoning_bytes": 107.538, "flops_per_problem": 6250929537.024}, "runmax.rev12": {"acc": 1.0, "reasoning_bytes": 83.0, "flops_per_problem": 4857937920.0}, "runmax.rev14": {"acc": 0.998, "reasoning_bytes": 94.988, "flops_per_problem": 5475690971.136}, "runmax.rev16": {"acc": 0.984, "reasoning_bytes": 101.968, "flops_per_problem": 5872275357.696}, "runmax.sumd12": {"acc": 1.0, "reasoning_bytes": 86.0, "flops_per_problem": 5034590208.0}, "runmax.sumd14": {"acc": 0.998, "reasoning_bytes": 97.892, "flops_per_problem": 5648103604.224}, "runmax.sumd16": {"acc": 0.978, "reasoning_bytes": 104.684, "flops_per_problem": 6036385333.248}, "runmax.inc12": {"acc": 1.0, "reasoning_bytes": 87.0, "flops_per_problem": 5065769336.832}, "runmax.inc14": {"acc": 0.998, "reasoning_bytes": 98.964, "flops_per_problem": 5685818867.712}, "runmax.inc16": {"acc": 0.984, "reasoning_bytes": 106.15, "flops_per_problem": 6093444022.272}, "count.cmp16": {"acc": 0.994, "reasoning_bytes": 95.0, "flops_per_problem": 5697036288.0}, "count.cmp18": {"acc": 0.996, "reasoning_bytes": 104.67, "flops_per_problem": 6212507664.384}, "count.cmp20": {"acc": 0.974, "reasoning_bytes": 111.886, "flops_per_problem": 6619602862.08}, "count.runmax16": {"acc": 1.0, "reasoning_bytes": 96.0, "flops_per_problem": 5785362432.0}, "count.runmax18": {"acc": 0.982, "reasoning_bytes": 105.64, "flops_per_problem": 6299420590.08}, "count.runmax20": {"acc": 0.846, "reasoning_bytes": 112.466, "flops_per_problem": 6689203863.552}, "count.count16": {"acc": 1.0, "reasoning_bytes": 95.99, "flops_per_problem": 5829083873.28}, "count.count18": {"acc": 1.0, "reasoning_bytes": 105.86, "flops_per_problem": 6353299537.92}, "count.count20": {"acc": 0.996, "reasoning_bytes": 113.24, "flops_per_problem": 6767549153.28}, "count.find16": {"acc": 1.0, "reasoning_bytes": 95.99, "flops_per_problem": 5784920801.28}, "count.find18": {"acc": 0.992, "reasoning_bytes": 105.64, "flops_per_problem": 6299420590.08}, "count.find20": {"acc": 0.968, "reasoning_bytes": 112.996, "flops_per_problem": 6712610291.712}, "count.rev16": {"acc": 1.0, "reasoning_bytes": 91.0, "flops_per_problem": 5432057856.0}, "count.rev18": {"acc": 0.988, "reasoning_bytes": 100.76, "flops_per_problem": 5951415582.72}, "count.rev20": {"acc": 0.856, "reasoning_bytes": 107.982, "flops_per_problem": 6358687432.704}, "count.sumd16": {"acc": 1.0, "reasoning_bytes": 94.0, "flops_per_problem": 5608710144.0}, "count.sumd18": {"acc": 0.988, "reasoning_bytes": 103.79, "flops_per_problem": 6129392762.88}, "count.sumd20": {"acc": 0.844, "reasoning_bytes": 110.952, "flops_per_problem": 6534103154.688}, "count.inc16": {"acc": 1.0, "reasoning_bytes": 95.0, "flops_per_problem": 5608798470.144}, "count.inc18": {"acc": 0.994, "reasoning_bytes": 104.8, "flops_per_problem": 6130011045.888}, "count.inc20": {"acc": 0.866, "reasoning_bytes": 112.296, "flops_per_problem": 6549206925.312}, "parity.cmp16": {"acc": 1.0, "reasoning_bytes": 110.976, "flops_per_problem": 6358422454.272}, "parity.cmp20": {"acc": 0.996, "reasoning_bytes": 134.004, "flops_per_problem": 7552061964.288}, "parity.cmp24": {"acc": 0.93, "reasoning_bytes": 151.416, "flops_per_problem": 8497681661.952}, "parity.runmax16": {"acc": 1.0, "reasoning_bytes": 112.0, "flops_per_problem": 6447808512.0}, "parity.runmax20": {"acc": 0.934, "reasoning_bytes": 135.232, "flops_per_problem": 7650457288.704}, "parity.runmax24": {"acc": 0.348, "reasoning_bytes": 152.012, "flops_per_problem": 8568165924.864}, "parity.count16": {"acc": 1.0, "reasoning_bytes": 112.0, "flops_per_problem": 6491971584.0}, "parity.count20": {"acc": 1.0, "reasoning_bytes": 135.256, "flops_per_problem": 7695680274.432}, "parity.count24": {"acc": 1.0, "reasoning_bytes": 152.32, "flops_per_problem": 8625931223.04}, "parity.find16": {"acc": 1.0, "reasoning_bytes": 111.964, "flops_per_problem": 6446218641.408}, "parity.find20": {"acc": 0.968, "reasoning_bytes": 135.388, "flops_per_problem": 7657346727.936}, "parity.find24": {"acc": 0.862, "reasoning_bytes": 153.028, "flops_per_problem": 8613035606.016}, "parity.rev16": {"acc": 1.0, "reasoning_bytes": 107.0, "flops_per_problem": 6094503936.0}, "parity.rev20": {"acc": 0.938, "reasoning_bytes": 130.436, "flops_per_problem": 7306161979.392}, "parity.rev24": {"acc": 0.422, "reasoning_bytes": 148.124, "flops_per_problem": 8263970684.928}, "parity.sumd16": {"acc": 1.0, "reasoning_bytes": 110.0, "flops_per_problem": 6271156224.0}, "parity.sumd20": {"acc": 0.912, "reasoning_bytes": 133.124, "flops_per_problem": 7469035388.928}, "parity.sumd24": {"acc": 0.434, "reasoning_bytes": 150.824, "flops_per_problem": 8427374051.328}, "parity.inc16": {"acc": 1.0, "reasoning_bytes": 111.0, "flops_per_problem": 6271156224.0}, "parity.inc20": {"acc": 0.926, "reasoning_bytes": 134.136, "flops_per_problem": 7469565345.792}, "parity.inc24": {"acc": 0.456, "reasoning_bytes": 152.292, "flops_per_problem": 8448042369.024}, "sumd.cmp6": {"acc": 1.0, "reasoning_bytes": 62.088, "flops_per_problem": 3713407746.048}, "sumd.cmp7": {"acc": 1.0, "reasoning_bytes": 70.084, "flops_per_problem": 4110875394.048}, "sumd.cmp8": {"acc": 0.996, "reasoning_bytes": 78.004, "flops_per_problem": 4504721670.144}, "sumd.runmax6": {"acc": 1.0, "reasoning_bytes": 67.542, "flops_per_problem": 3954449793.024}, "sumd.runmax7": {"acc": 1.0, "reasoning_bytes": 75.48, "flops_per_problem": 4349179330.56}, "sumd.runmax8": {"acc": 0.988, "reasoning_bytes": 83.438, "flops_per_problem": 4744792129.536}, "sumd.count6": {"acc": 1.0, "reasoning_bytes": 66.428, "flops_per_problem": 3949415202.816}, "sumd.count7": {"acc": 1.0, "reasoning_bytes": 74.46, "flops_per_problem": 4348296069.12}, "sumd.count8": {"acc": 1.0, "reasoning_bytes": 82.588, "flops_per_problem": 4751416590.336}, "sumd.find6": {"acc": 1.0, "reasoning_bytes": 66.24, "flops_per_problem": 3896949473.28}, "sumd.find7": {"acc": 1.0, "reasoning_bytes": 74.428, "flops_per_problem": 4302719778.816}, "sumd.find8": {"acc": 0.998, "reasoning_bytes": 82.296, "flops_per_problem": 4694357901.312}, "sumd.rev6": {"acc": 1.0, "reasoning_bytes": 61.544, "flops_per_problem": 3601145217.024}, "sumd.rev7": {"acc": 1.0, "reasoning_bytes": 69.48, "flops_per_problem": 3995874754.56}, "sumd.rev8": {"acc": 0.998, "reasoning_bytes": 77.552, "flops_per_problem": 4396433817.6}, "sumd.sumd6": {"acc": 1.0, "reasoning_bytes": 65.744, "flops_per_problem": 3795639386.112}, "sumd.sumd7": {"acc": 1.0, "reasoning_bytes": 73.738, "flops_per_problem": 4195315187.712}, "sumd.sumd8": {"acc": 0.994, "reasoning_bytes": 81.836, "flops_per_problem": 4598347382.784}, "sumd.inc6": {"acc": 1.0, "reasoning_bytes": 68.538, "flops_per_problem": 3910021742.592}, "sumd.inc7": {"acc": 0.998, "reasoning_bytes": 76.48, "flops_per_problem": 4305016258.56}, "sumd.inc8": {"acc": 0.936, "reasoning_bytes": 84.498, "flops_per_problem": 4703455494.144}, "inc.cmp6": {"acc": 0.996, "reasoning_bytes": 78.624, "flops_per_problem": 4576354172.928}, "inc.cmp7": {"acc": 0.998, "reasoning_bytes": 89.806, "flops_per_problem": 5158511788.032}, "inc.cmp8": {"acc": 0.986, "reasoning_bytes": 100.754, "flops_per_problem": 5730335244.288}, "inc.runmax6": {"acc": 1.0, "reasoning_bytes": 104.0, "flops_per_problem": 5520384000.0}, "inc.runmax7": {"acc": 1.0, "reasoning_bytes": 120.0, "flops_per_problem": 6271156224.0}, "inc.runmax8": {"acc": 0.998, "reasoning_bytes": 135.952, "flops_per_problem": 7019808620.544}, "inc.count6": {"acc": 0.988, "reasoning_bytes": 99.0, "flops_per_problem": 5343731712.0}, "inc.count7": {"acc": 0.986, "reasoning_bytes": 114.0, "flops_per_problem": 6050340864.0}, "inc.count8": {"acc": 0.936, "reasoning_bytes": 128.996, "flops_per_problem": 6756773363.712}, "inc.find6": {"acc": 1.0, "reasoning_bytes": 91.88, "flops_per_problem": 4985127567.36}, "inc.find7": {"acc": 1.0, "reasoning_bytes": 102.516, "flops_per_problem": 5499009073.152}, "inc.find8": {"acc": 0.876, "reasoning_bytes": 116.92, "flops_per_problem": 6179208708.096}, "inc.rev6": {"acc": 1.0, "reasoning_bytes": 104.0, "flops_per_problem": 5608710144.0}, "inc.rev7": {"acc": 1.0, "reasoning_bytes": 124.0, "flops_per_problem": 6580297728.0}, "inc.rev8": {"acc": 0.97, "reasoning_bytes": 145.0, "flops_per_problem": 7596048384.0}, "inc.sumd6": {"acc": 1.0, "reasoning_bytes": 109.558, "flops_per_problem": 5721590956.032}, "inc.sumd7": {"acc": 0.988, "reasoning_bytes": 127.466, "flops_per_problem": 6556626321.408}, "inc.sumd8": {"acc": 0.872, "reasoning_bytes": 145.436, "flops_per_problem": 7394488123.392}, "inc.inc6": {"acc": 1.0, "reasoning_bytes": 113.0, "flops_per_problem": 6006177792.0}, "inc.inc7": {"acc": 0.998, "reasoning_bytes": 131.0, "flops_per_problem": 6889527558.144}, "inc.inc8": {"acc": 0.97, "reasoning_bytes": 148.964, "flops_per_problem": 7772082388.992}, "x.add.cmp8": {"acc": 0.848, "reasoning_bytes": 150.156, "flops_per_problem": 8425254223.872}, "x.add.cmp9": {"acc": 0.748, "reasoning_bytes": 166.056, "flops_per_problem": 9265147527.168}, "x.add.cmp10": {"acc": 0.666, "reasoning_bytes": 178.382, "flops_per_problem": 9972109983.744}, "x.add.runmax8": {"acc": 1.0, "reasoning_bytes": 189.848, "flops_per_problem": 9885815341.056}, "x.add.runmax9": {"acc": 0.998, "reasoning_bytes": 209.808, "flops_per_problem": 10855636402.176}, "x.add.runmax10": {"acc": 0.952, "reasoning_bytes": 228.586, "flops_per_problem": 11772726755.328}, "x.add.count8": {"acc": 0.972, "reasoning_bytes": 180.948, "flops_per_problem": 9536927072.256}, "x.add.count9": {"acc": 0.934, "reasoning_bytes": 199.998, "flops_per_problem": 10466559737.856}, "x.add.count10": {"acc": 0.756, "reasoning_bytes": 216.396, "flops_per_problem": 11279071936.512}, "x.add.find8": {"acc": 0.914, "reasoning_bytes": 165.548, "flops_per_problem": 8812652691.456}, "x.add.find9": {"acc": 0.848, "reasoning_bytes": 182.464, "flops_per_problem": 9648306339.84}, "x.add.find10": {"acc": 0.754, "reasoning_bytes": 199.712, "flops_per_problem": 10499505389.568}, "x.add.rev8": {"acc": 0.994, "reasoning_bytes": 199.484, "flops_per_problem": 10514785812.48}, "x.add.rev9": {"acc": 0.812, "reasoning_bytes": 225.356, "flops_per_problem": 11787212242.944}, "x.add.rev10": {"acc": 0.116, "reasoning_bytes": 244.31, "flops_per_problem": 12728150654.976}, "x.add.sumd8": {"acc": 0.966, "reasoning_bytes": 197.284, "flops_per_problem": 10170048872.448}, "x.add.sumd9": {"acc": 0.78, "reasoning_bytes": 218.554, "flops_per_problem": 11197723557.888}, "x.add.sumd10": {"acc": 0.46, "reasoning_bytes": 240.028, "flops_per_problem": 12234319183.872}, "x.add.inc8": {"acc": 0.998, "reasoning_bytes": 201.06, "flops_per_problem": 10584386813.952}, "x.add.inc9": {"acc": 0.896, "reasoning_bytes": 223.1, "flops_per_problem": 11693939834.88}, "x.add.inc10": {"acc": 0.266, "reasoning_bytes": 241.916, "flops_per_problem": 12670650335.232}, "x.runmax.cmp12": {"acc": 1.0, "reasoning_bytes": 118.0, "flops_per_problem": 6580297728.0}, "x.runmax.cmp14": {"acc": 1.0, "reasoning_bytes": 131.964, "flops_per_problem": 7285317009.408}, "x.runmax.cmp16": {"acc": 1.0, "reasoning_bytes": 144.95, "flops_per_problem": 7947144806.4}, "x.runmax.runmax12": {"acc": 1.0, "reasoning_bytes": 122.0, "flops_per_problem": 6801113088.0}, "x.runmax.runmax14": {"acc": 1.0, "reasoning_bytes": 136.074, "flops_per_problem": 7510990307.328}, "x.runmax.runmax16": {"acc": 0.994, "reasoning_bytes": 149.45, "flops_per_problem": 8190041702.4}, "x.runmax.count12": {"acc": 0.876, "reasoning_bytes": 120.496, "flops_per_problem": 6778854899.712}, "x.runmax.count14": {"acc": 0.786, "reasoning_bytes": 134.052, "flops_per_problem": 7465855647.744}, "x.runmax.count16": {"acc": 0.684, "reasoning_bytes": 145.654, "flops_per_problem": 8066561753.088}, "x.runmax.find12": {"acc": 1.0, "reasoning_bytes": 120.0, "flops_per_problem": 6712786944.0}, "x.runmax.find14": {"acc": 1.0, "reasoning_bytes": 134.094, "flops_per_problem": 7423547424.768}, "x.runmax.find16": {"acc": 0.998, "reasoning_bytes": 147.522, "flops_per_problem": 8105248604.16}, "x.runmax.rev12": {"acc": 1.0, "reasoning_bytes": 114.0, "flops_per_problem": 6315319296.0}, "x.runmax.rev14": {"acc": 1.0, "reasoning_bytes": 128.044, "flops_per_problem": 7023871623.168}, "x.runmax.rev16": {"acc": 0.994, "reasoning_bytes": 142.232, "flops_per_problem": 7738783432.704}, "x.runmax.sumd12": {"acc": 1.0, "reasoning_bytes": 118.0, "flops_per_problem": 6536134656.0}, "x.runmax.sumd14": {"acc": 1.0, "reasoning_bytes": 132.072, "flops_per_problem": 7245923549.184}, "x.runmax.sumd16": {"acc": 0.99, "reasoning_bytes": 144.622, "flops_per_problem": 7888496246.784}, "x.runmax.inc12": {"acc": 1.0, "reasoning_bytes": 118.0, "flops_per_problem": 6523150712.832}, "x.runmax.inc14": {"acc": 0.998, "reasoning_bytes": 132.06, "flops_per_problem": 7235677716.48}, "x.runmax.inc16": {"acc": 0.99, "reasoning_bytes": 144.766, "flops_per_problem": 7887348006.912}, "x.count.cmp16": {"acc": 0.996, "reasoning_bytes": 129.968, "flops_per_problem": 7329656733.696}, "x.count.cmp18": {"acc": 0.996, "reasoning_bytes": 141.812, "flops_per_problem": 7941138628.608}, "x.count.cmp20": {"acc": 0.972, "reasoning_bytes": 152.178, "flops_per_problem": 8487347503.104}, "x.count.runmax16": {"acc": 0.996, "reasoning_bytes": 133.992, "flops_per_problem": 7551532007.424}, "x.count.runmax18": {"acc": 0.982, "reasoning_bytes": 145.804, "flops_per_problem": 8161512357.888}, "x.count.runmax20": {"acc": 0.9, "reasoning_bytes": 156.55, "flops_per_problem": 8724326547.456}, "x.count.count16": {"acc": 0.998, "reasoning_bytes": 132.936, "flops_per_problem": 7549058875.392}, "x.count.count18": {"acc": 1.0, "reasoning_bytes": 144.82, "flops_per_problem": 8162218967.04}, "x.count.count20": {"acc": 0.994, "reasoning_bytes": 155.22, "flops_per_problem": 8709841059.84}, "x.count.find16": {"acc": 0.994, "reasoning_bytes": 131.994, "flops_per_problem": 7463294189.568}, "x.count.find18": {"acc": 0.98, "reasoning_bytes": 143.796, "flops_per_problem": 8072832909.312}, "x.count.find20": {"acc": 0.952, "reasoning_bytes": 154.576, "flops_per_problem": 8637148643.328}, "x.count.rev16": {"acc": 0.998, "reasoning_bytes": 125.982, "flops_per_problem": 7065296584.704}, "x.count.rev18": {"acc": 0.99, "reasoning_bytes": 137.84, "flops_per_problem": 7677308436.48}, "x.count.rev20": {"acc": 0.928, "reasoning_bytes": 148.926, "flops_per_problem": 8255226396.672}, "x.count.sumd16": {"acc": 0.998, "reasoning_bytes": 129.98, "flops_per_problem": 7286023618.56}, "x.count.sumd18": {"acc": 0.988, "reasoning_bytes": 141.846, "flops_per_problem": 7898388774.912}, "x.count.sumd20": {"acc": 0.906, "reasoning_bytes": 152.758, "flops_per_problem": 8468799012.864}, "x.count.inc16": {"acc": 0.998, "reasoning_bytes": 129.99, "flops_per_problem": 7242390503.424}, "x.count.inc18": {"acc": 0.994, "reasoning_bytes": 141.934, "flops_per_problem": 7858200379.392}, "x.count.inc20": {"acc": 0.936, "reasoning_bytes": 153.072, "flops_per_problem": 8438414819.328}, "x.parity.cmp24": {"acc": 0.996, "reasoning_bytes": 201.97, "flops_per_problem": 10818627747.84}, "x.parity.cmp28": {"acc": 0.982, "reasoning_bytes": 229.662, "flops_per_problem": 12218243825.664}, "x.parity.cmp32": {"acc": 0.926, "reasoning_bytes": 250.74, "flops_per_problem": 13325765345.28}, "x.parity.runmax24": {"acc": 1.0, "reasoning_bytes": 206.032, "flops_per_problem": 11042181218.304}, "x.parity.runmax28": {"acc": 0.828, "reasoning_bytes": 233.536, "flops_per_problem": 12433406312.448}, "x.parity.runmax32": {"acc": 0.36, "reasoning_bytes": 254.866, "flops_per_problem": 13552145252.352}, "x.parity.count24": {"acc": 0.994, "reasoning_bytes": 204.98, "flops_per_problem": 11039884738.56}, "x.parity.count28": {"acc": 1.0, "reasoning_bytes": 232.28, "flops_per_problem": 12422188892.16}, "x.parity.count32": {"acc": 0.982, "reasoning_bytes": 254.264, "flops_per_problem": 13570428764.16}, "x.parity.find24": {"acc": 1.0, "reasoning_bytes": 204.03, "flops_per_problem": 10953766748.16}, "x.parity.find28": {"acc": 0.958, "reasoning_bytes": 231.588, "flops_per_problem": 12347464974.336}, "x.parity.find32": {"acc": 0.846, "reasoning_bytes": 252.77, "flops_per_problem": 13459226148.864}, "x.parity.rev24": {"acc": 0.778, "reasoning_bytes": 198.142, "flops_per_problem": 10560715407.36}, "x.parity.rev28": {"acc": 0.652, "reasoning_bytes": 226.804, "flops_per_problem": 12012355584.0}, "x.parity.rev32": {"acc": 0.454, "reasoning_bytes": 252.948, "flops_per_problem": 13338131005.44}, "x.parity.sumd24": {"acc": 0.988, "reasoning_bytes": 202.04, "flops_per_problem": 10777556090.88}, "x.parity.sumd28": {"acc": 0.854, "reasoning_bytes": 230.232, "flops_per_problem": 12199253704.704}, "x.parity.sumd32": {"acc": 0.46, "reasoning_bytes": 252.536, "flops_per_problem": 13360565846.016}, "x.parity.inc24": {"acc": 0.98, "reasoning_bytes": 202.14, "flops_per_problem": 10737809326.08}, "x.parity.inc28": {"acc": 0.826, "reasoning_bytes": 230.02, "flops_per_problem": 12145728061.44}, "x.parity.inc32": {"acc": 0.454, "reasoning_bytes": 252.07, "flops_per_problem": 13296176087.04}, "x.sumd.cmp8": {"acc": 1.0, "reasoning_bytes": 101.096, "flops_per_problem": 5612861472.768}, "x.sumd.cmp9": {"acc": 1.0, "reasoning_bytes": 110.252, "flops_per_problem": 6061469958.144}, "x.sumd.cmp10": {"acc": 0.994, "reasoning_bytes": 119.168, "flops_per_problem": 6499390980.096}, "x.sumd.runmax8": {"acc": 1.0, "reasoning_bytes": 109.55, "flops_per_problem": 5986304409.6}, "x.sumd.runmax9": {"acc": 0.99, "reasoning_bytes": 118.432, "flops_per_problem": 6422723887.104}, "x.sumd.runmax10": {"acc": 0.966, "reasoning_bytes": 127.534, "flops_per_problem": 6868859240.448}, "x.sumd.count8": {"acc": 0.998, "reasoning_bytes": 107.588, "flops_per_problem": 5943819534.336}, "x.sumd.count9": {"acc": 0.994, "reasoning_bytes": 116.33, "flops_per_problem": 6374144507.904}, "x.sumd.count10": {"acc": 0.968, "reasoning_bytes": 125.558, "flops_per_problem": 6825756082.176}, "x.sumd.find8": {"acc": 1.0, "reasoning_bytes": 106.33, "flops_per_problem": 5844099317.76}, "x.sumd.find9": {"acc": 0.998, "reasoning_bytes": 115.448, "flops_per_problem": 6290941280.256}, "x.sumd.find10": {"acc": 0.972, "reasoning_bytes": 124.378, "flops_per_problem": 6729480585.216}, "x.sumd.rev8": {"acc": 1.0, "reasoning_bytes": 100.552, "flops_per_problem": 5500510617.6}, "x.sumd.rev9": {"acc": 0.978, "reasoning_bytes": 109.432, "flops_per_problem": 5936930095.104}, "x.sumd.rev10": {"acc": 0.936, "reasoning_bytes": 118.542, "flops_per_problem": 6383683731.456}, "x.sumd.sumd8": {"acc": 0.988, "reasoning_bytes": 105.838, "flops_per_problem": 5746763907.072}, "x.sumd.sumd9": {"acc": 0.948, "reasoning_bytes": 114.8, "flops_per_problem": 6190337802.24}, "x.sumd.sumd10": {"acc": 0.88, "reasoning_bytes": 123.886, "flops_per_problem": 6637886373.888}, "x.sumd.inc8": {"acc": 0.516, "reasoning_bytes": 107.55, "flops_per_problem": 5810712035.328}, "x.sumd.inc9": {"acc": 0.474, "reasoning_bytes": 116.378, "flops_per_problem": 6248809709.568}, "x.sumd.inc10": {"acc": 0.34, "reasoning_bytes": 125.48, "flops_per_problem": 6697859825.664}, "x.inc.cmp8": {"acc": 0.998, "reasoning_bytes": 121.792, "flops_per_problem": 6747764097.024}, "x.inc.cmp9": {"acc": 0.996, "reasoning_bytes": 133.744, "flops_per_problem": 7363927277.568}, "x.inc.cmp10": {"acc": 0.968, "reasoning_bytes": 145.532, "flops_per_problem": 7985655005.184}, "x.inc.runmax8": {"acc": 0.998, "reasoning_bytes": 159.988, "flops_per_problem": 8169638363.136}, "x.inc.runmax9": {"acc": 1.0, "reasoning_bytes": 176.962, "flops_per_problem": 8963425419.264}, "x.inc.runmax10": {"acc": 1.0, "reasoning_bytes": 193.308, "flops_per_problem": 9729478066.176}, "x.inc.count8": {"acc": 0.986, "reasoning_bytes": 152.0, "flops_per_problem": 7861026816.0}, "x.inc.count9": {"acc": 0.972, "reasoning_bytes": 167.86, "flops_per_problem": 8605616209.92}, "x.inc.count10": {"acc": 0.914, "reasoning_bytes": 182.598, "flops_per_problem": 9300831289.344}, "x.inc.find8": {"acc": 0.96, "reasoning_bytes": 138.536, "flops_per_problem": 7222252142.592}, "x.inc.find9": {"acc": 0.862, "reasoning_bytes": 153.408, "flops_per_problem": 7923208421.376}, "x.inc.find10": {"acc": 0.808, "reasoning_bytes": 164.936, "flops_per_problem": 8476483387.392}, "x.inc.rev8": {"acc": 1.0, "reasoning_bytes": 166.0, "flops_per_problem": 8611799040.0}, "x.inc.rev9": {"acc": 0.968, "reasoning_bytes": 188.918, "flops_per_problem": 9711901163.52}, "x.inc.rev10": {"acc": 0.424, "reasoning_bytes": 208.354, "flops_per_problem": 10644713570.304}, "x.inc.sumd8": {"acc": 0.936, "reasoning_bytes": 167.476, "flops_per_problem": 8456168374.272}, "x.inc.sumd9": {"acc": 0.634, "reasoning_bytes": 185.624, "flops_per_problem": 9301626224.64}, "x.inc.sumd10": {"acc": 0.142, "reasoning_bytes": 202.368, "flops_per_problem": 10085520752.64}, "x.inc.inc8": {"acc": 0.99, "reasoning_bytes": 170.0, "flops_per_problem": 8788892958.72}, "x.inc.inc9": {"acc": 0.824, "reasoning_bytes": 188.924, "flops_per_problem": 9718702276.608}, "x.inc.inc10": {"acc": 0.22, "reasoning_bytes": 206.016, "flops_per_problem": 10580235485.184}} diff --git a/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/summary.json b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..fc72a8785435fac718a3f9e48a4dcc458a5352cb --- /dev/null +++ b/5m-s0-bpe4096-step-bytes3-v11-nox.add.cmp-nox.runmax.count-nox.count.find-nox.parity.rev-nox.sumd.inc-nox.inc.sumd-none-docmask-c5090/summary.json @@ -0,0 +1,1405 @@ +{ + "params": 5787648, + "tokens": 399998976, + "train_flops": 17665183577014272, + "wall_s": 1227.6587660312653, + "peak_mem_gb": 2.254330368, + "final_val_loss": 1.806345921754837, + "final_val_bpb": 0.651158358435283, + "bytes_seen": 1600839241.8284984, + "add8": { + "acc": 1.0, + "reasoning_bytes": 79.0, + "flops_per_problem": 4928775487.488 + }, + "add9": { + "acc": 0.952, + "reasoning_bytes": 89.0, + "flops_per_problem": 5503160401.92 + }, + "add10": { + "acc": 0.382, + "reasoning_bytes": 99.0, + "flops_per_problem": 6077810294.784 + }, + "reach14": { + "acc": 0.586, + "reasoning_bytes": 7.6, + "flops_per_problem": 4619457331.2 + }, + "reach16": { + "acc": 0.566, + "reasoning_bytes": 7.592, + "flops_per_problem": 5149237542.912 + }, + "reach18": { + "acc": 0.482, + "reasoning_bytes": 7.558, + "flops_per_problem": 5677692862.464 + }, + "parity24": { + "acc": 0.852, + "reasoning_bytes": 142.112, + "flops_per_problem": 7821810008.064 + }, + "parity28": { + "acc": 0.454, + "reasoning_bytes": 161.044, + "flops_per_problem": 8834557575.168 + }, + "parity32": { + "acc": 0.434, + "reasoning_bytes": 181.224, + "flops_per_problem": 9902420656.128 + }, + "cmp8": { + "acc": 0.998, + "reasoning_bytes": 5.836, + "flops_per_problem": 1362077466.624 + }, + "cmp9": { + "acc": 1.0, + "reasoning_bytes": 5.708, + "flops_per_problem": 1444485758.976 + }, + "cmp10": { + "acc": 1.0, + "reasoning_bytes": 5.756, + "flops_per_problem": 1534931730.432 + }, + "runmax12": { + "acc": 1.0, + "reasoning_bytes": 71.0, + "flops_per_problem": 4151328768.0 + }, + "runmax14": { + "acc": 1.0, + "reasoning_bytes": 83.0, + "flops_per_problem": 4769611776.0 + }, + "runmax16": { + "acc": 0.992, + "reasoning_bytes": 92.372, + "flops_per_problem": 5271834230.784 + }, + "count16": { + "acc": 0.996, + "reasoning_bytes": 79.006, + "flops_per_problem": 4725625356.288 + }, + "count18": { + "acc": 0.99, + "reasoning_bytes": 88.74, + "flops_per_problem": 5243923169.28 + }, + "count20": { + "acc": 0.91, + "reasoning_bytes": 96.806, + "flops_per_problem": 5688468652.032 + }, + "brackets16": { + "acc": 0.962, + "reasoning_bytes": 52.004, + "flops_per_problem": 3577385484.288 + }, + "brackets18": { + "acc": 0.968, + "reasoning_bytes": 53.426, + "flops_per_problem": 3728511516.672 + }, + "brackets20": { + "acc": 0.976, + "reasoning_bytes": 59.49, + "flops_per_problem": 4084642529.28 + }, + "sumd8": { + "acc": 0.994, + "reasoning_bytes": 58.566, + "flops_per_problem": 3381301444.608 + }, + "sumd9": { + "acc": 0.964, + "reasoning_bytes": 66.426, + "flops_per_problem": 3772674588.672 + }, + "sumd10": { + "acc": 0.876, + "reasoning_bytes": 74.364, + "flops_per_problem": 4167315800.064 + }, + "inc8": { + "acc": 1.0, + "reasoning_bytes": 63.0, + "flops_per_problem": 3798024192.0 + }, + "inc9": { + "acc": 0.988, + "reasoning_bytes": 71.0, + "flops_per_problem": 4239654912.0 + }, + "inc10": { + "acc": 0.452, + "reasoning_bytes": 76.248, + "flops_per_problem": 4544556761.088 + }, + "add.cmp6": { + "acc": 0.998, + "reasoning_bytes": 92.622, + "flops_per_problem": 5532043051.008 + }, + "add.cmp7": { + "acc": 1.0, + "reasoning_bytes": 105.548, + "flops_per_problem": 6232646025.216 + }, + "add.cmp8": { + "acc": 0.98, + "reasoning_bytes": 118.842, + "flops_per_problem": 6953740664.832 + }, + "add.runmax6": { + "acc": 1.0, + "reasoning_bytes": 120.832, + "flops_per_problem": 6572878331.904 + }, + "add.runmax7": { + "acc": 1.0, + "reasoning_bytes": 138.848, + "flops_per_problem": 7456846381.056 + }, + "add.runmax8": { + "acc": 0.976, + "reasoning_bytes": 156.832, + "flops_per_problem": 8339401211.904 + }, + "add.count6": { + "acc": 0.994, + "reasoning_bytes": 115.648, + "flops_per_problem": 6388100038.656 + }, + "add.count7": { + "acc": 0.99, + "reasoning_bytes": 132.088, + "flops_per_problem": 7202467086.336 + }, + "add.count8": { + "acc": 0.746, + "reasoning_bytes": 148.904, + "flops_per_problem": 8033439449.088 + }, + "add.find6": { + "acc": 0.998, + "reasoning_bytes": 107.696, + "flops_per_problem": 5992752218.112 + }, + "add.find7": { + "acc": 0.972, + "reasoning_bytes": 121.2, + "flops_per_problem": 6677456486.4 + }, + "add.find8": { + "acc": 0.818, + "reasoning_bytes": 135.224, + "flops_per_problem": 7385125552.128 + }, + "add.rev6": { + "acc": 1.0, + "reasoning_bytes": 123.248, + "flops_per_problem": 6794576953.344 + }, + "add.rev7": { + "acc": 1.0, + "reasoning_bytes": 145.878, + "flops_per_problem": 7926564814.848 + }, + "add.rev8": { + "acc": 0.852, + "reasoning_bytes": 169.456, + "flops_per_problem": 9100242616.32 + }, + "add.sumd6": { + "acc": 1.0, + "reasoning_bytes": 126.476, + "flops_per_problem": 6777794985.984 + }, + "add.sumd7": { + "acc": 0.99, + "reasoning_bytes": 146.33, + "flops_per_problem": 7743023087.616 + }, + "add.sumd8": { + "acc": 0.806, + "reasoning_bytes": 166.244, + "flops_per_problem": 8710900973.568 + }, + "add.inc6": { + "acc": 1.0, + "reasoning_bytes": 131.04, + "flops_per_problem": 7138695610.368 + }, + "add.inc7": { + "acc": 1.0, + "reasoning_bytes": 151.06, + "flops_per_problem": 8155417853.952 + }, + "add.inc8": { + "acc": 0.93, + "reasoning_bytes": 171.04, + "flops_per_problem": 9171080183.808 + }, + "runmax.cmp12": { + "acc": 1.0, + "reasoning_bytes": 87.0, + "flops_per_problem": 5122916352.0 + }, + "runmax.cmp14": { + "acc": 1.0, + "reasoning_bytes": 98.968, + "flops_per_problem": 5739786141.696 + }, + "runmax.cmp16": { + "acc": 0.994, + "reasoning_bytes": 107.1, + "flops_per_problem": 6187246387.2 + }, + "runmax.runmax12": { + "acc": 1.0, + "reasoning_bytes": 88.0, + "flops_per_problem": 5211242496.0 + }, + "runmax.runmax14": { + "acc": 0.998, + "reasoning_bytes": 99.916, + "flops_per_problem": 5825815805.952 + }, + "runmax.runmax16": { + "acc": 0.976, + "reasoning_bytes": 106.3, + "flops_per_problem": 6196079001.6 + }, + "runmax.count12": { + "acc": 1.0, + "reasoning_bytes": 88.0, + "flops_per_problem": 5255405568.0 + }, + "runmax.count14": { + "acc": 1.0, + "reasoning_bytes": 99.976, + "flops_per_problem": 5872628662.272 + }, + "runmax.count16": { + "acc": 0.992, + "reasoning_bytes": 107.23, + "flops_per_problem": 6281313730.56 + }, + "runmax.find12": { + "acc": 1.0, + "reasoning_bytes": 88.0, + "flops_per_problem": 5211242496.0 + }, + "runmax.find14": { + "acc": 1.0, + "reasoning_bytes": 99.956, + "flops_per_problem": 5827582328.832 + }, + "runmax.find16": { + "acc": 0.994, + "reasoning_bytes": 107.538, + "flops_per_problem": 6250929537.024 + }, + "runmax.rev12": { + "acc": 1.0, + "reasoning_bytes": 83.0, + "flops_per_problem": 4857937920.0 + }, + "runmax.rev14": { + "acc": 0.998, + "reasoning_bytes": 94.988, + "flops_per_problem": 5475690971.136 + }, + "runmax.rev16": { + "acc": 0.984, + "reasoning_bytes": 101.968, + "flops_per_problem": 5872275357.696 + }, + "runmax.sumd12": { + "acc": 1.0, + "reasoning_bytes": 86.0, + "flops_per_problem": 5034590208.0 + }, + "runmax.sumd14": { + "acc": 0.998, + "reasoning_bytes": 97.892, + "flops_per_problem": 5648103604.224 + }, + "runmax.sumd16": { + "acc": 0.978, + "reasoning_bytes": 104.684, + "flops_per_problem": 6036385333.248 + }, + "runmax.inc12": { + "acc": 1.0, + "reasoning_bytes": 87.0, + "flops_per_problem": 5065769336.832 + }, + "runmax.inc14": { + "acc": 0.998, + "reasoning_bytes": 98.964, + "flops_per_problem": 5685818867.712 + }, + "runmax.inc16": { + "acc": 0.984, + "reasoning_bytes": 106.15, + "flops_per_problem": 6093444022.272 + }, + "count.cmp16": { + "acc": 0.994, + "reasoning_bytes": 95.0, + "flops_per_problem": 5697036288.0 + }, + "count.cmp18": { + "acc": 0.996, + "reasoning_bytes": 104.67, + "flops_per_problem": 6212507664.384 + }, + "count.cmp20": { + "acc": 0.974, + "reasoning_bytes": 111.886, + "flops_per_problem": 6619602862.08 + }, + "count.runmax16": { + "acc": 1.0, + "reasoning_bytes": 96.0, + "flops_per_problem": 5785362432.0 + }, + "count.runmax18": { + "acc": 0.982, + "reasoning_bytes": 105.64, + "flops_per_problem": 6299420590.08 + }, + "count.runmax20": { + "acc": 0.846, + "reasoning_bytes": 112.466, + "flops_per_problem": 6689203863.552 + }, + "count.count16": { + "acc": 1.0, + "reasoning_bytes": 95.99, + "flops_per_problem": 5829083873.28 + }, + "count.count18": { + "acc": 1.0, + "reasoning_bytes": 105.86, + "flops_per_problem": 6353299537.92 + }, + "count.count20": { + "acc": 0.996, + "reasoning_bytes": 113.24, + "flops_per_problem": 6767549153.28 + }, + "count.find16": { + "acc": 1.0, + "reasoning_bytes": 95.99, + "flops_per_problem": 5784920801.28 + }, + "count.find18": { + "acc": 0.992, + "reasoning_bytes": 105.64, + "flops_per_problem": 6299420590.08 + }, + "count.find20": { + "acc": 0.968, + "reasoning_bytes": 112.996, + "flops_per_problem": 6712610291.712 + }, + "count.rev16": { + "acc": 1.0, + "reasoning_bytes": 91.0, + "flops_per_problem": 5432057856.0 + }, + "count.rev18": { + "acc": 0.988, + "reasoning_bytes": 100.76, + "flops_per_problem": 5951415582.72 + }, + "count.rev20": { + "acc": 0.856, + "reasoning_bytes": 107.982, + "flops_per_problem": 6358687432.704 + }, + "count.sumd16": { + "acc": 1.0, + "reasoning_bytes": 94.0, + "flops_per_problem": 5608710144.0 + }, + "count.sumd18": { + "acc": 0.988, + "reasoning_bytes": 103.79, + "flops_per_problem": 6129392762.88 + }, + "count.sumd20": { + "acc": 0.844, + "reasoning_bytes": 110.952, + "flops_per_problem": 6534103154.688 + }, + "count.inc16": { + "acc": 1.0, + "reasoning_bytes": 95.0, + "flops_per_problem": 5608798470.144 + }, + "count.inc18": { + "acc": 0.994, + "reasoning_bytes": 104.8, + "flops_per_problem": 6130011045.888 + }, + "count.inc20": { + "acc": 0.866, + "reasoning_bytes": 112.296, + "flops_per_problem": 6549206925.312 + }, + "parity.cmp16": { + "acc": 1.0, + "reasoning_bytes": 110.976, + "flops_per_problem": 6358422454.272 + }, + "parity.cmp20": { + "acc": 0.996, + "reasoning_bytes": 134.004, + "flops_per_problem": 7552061964.288 + }, + "parity.cmp24": { + "acc": 0.93, + "reasoning_bytes": 151.416, + "flops_per_problem": 8497681661.952 + }, + "parity.runmax16": { + "acc": 1.0, + "reasoning_bytes": 112.0, + "flops_per_problem": 6447808512.0 + }, + "parity.runmax20": { + "acc": 0.934, + "reasoning_bytes": 135.232, + "flops_per_problem": 7650457288.704 + }, + "parity.runmax24": { + "acc": 0.348, + "reasoning_bytes": 152.012, + "flops_per_problem": 8568165924.864 + }, + "parity.count16": { + "acc": 1.0, + "reasoning_bytes": 112.0, + "flops_per_problem": 6491971584.0 + }, + "parity.count20": { + "acc": 1.0, + "reasoning_bytes": 135.256, + "flops_per_problem": 7695680274.432 + }, + "parity.count24": { + "acc": 1.0, + "reasoning_bytes": 152.32, + "flops_per_problem": 8625931223.04 + }, + "parity.find16": { + "acc": 1.0, + "reasoning_bytes": 111.964, + "flops_per_problem": 6446218641.408 + }, + "parity.find20": { + "acc": 0.968, + "reasoning_bytes": 135.388, + "flops_per_problem": 7657346727.936 + }, + "parity.find24": { + "acc": 0.862, + "reasoning_bytes": 153.028, + "flops_per_problem": 8613035606.016 + }, + "parity.rev16": { + "acc": 1.0, + "reasoning_bytes": 107.0, + "flops_per_problem": 6094503936.0 + }, + "parity.rev20": { + "acc": 0.938, + "reasoning_bytes": 130.436, + "flops_per_problem": 7306161979.392 + }, + "parity.rev24": { + "acc": 0.422, + "reasoning_bytes": 148.124, + "flops_per_problem": 8263970684.928 + }, + "parity.sumd16": { + "acc": 1.0, + "reasoning_bytes": 110.0, + "flops_per_problem": 6271156224.0 + }, + "parity.sumd20": { + "acc": 0.912, + "reasoning_bytes": 133.124, + "flops_per_problem": 7469035388.928 + }, + "parity.sumd24": { + "acc": 0.434, + "reasoning_bytes": 150.824, + "flops_per_problem": 8427374051.328 + }, + "parity.inc16": { + "acc": 1.0, + "reasoning_bytes": 111.0, + "flops_per_problem": 6271156224.0 + }, + "parity.inc20": { + "acc": 0.926, + "reasoning_bytes": 134.136, + "flops_per_problem": 7469565345.792 + }, + "parity.inc24": { + "acc": 0.456, + "reasoning_bytes": 152.292, + "flops_per_problem": 8448042369.024 + }, + "sumd.cmp6": { + "acc": 1.0, + "reasoning_bytes": 62.088, + "flops_per_problem": 3713407746.048 + }, + "sumd.cmp7": { + "acc": 1.0, + "reasoning_bytes": 70.084, + "flops_per_problem": 4110875394.048 + }, + "sumd.cmp8": { + "acc": 0.996, + "reasoning_bytes": 78.004, + "flops_per_problem": 4504721670.144 + }, + "sumd.runmax6": { + "acc": 1.0, + "reasoning_bytes": 67.542, + "flops_per_problem": 3954449793.024 + }, + "sumd.runmax7": { + "acc": 1.0, + "reasoning_bytes": 75.48, + "flops_per_problem": 4349179330.56 + }, + "sumd.runmax8": { + "acc": 0.988, + "reasoning_bytes": 83.438, + "flops_per_problem": 4744792129.536 + }, + "sumd.count6": { + "acc": 1.0, + "reasoning_bytes": 66.428, + "flops_per_problem": 3949415202.816 + }, + "sumd.count7": { + "acc": 1.0, + "reasoning_bytes": 74.46, + "flops_per_problem": 4348296069.12 + }, + "sumd.count8": { + "acc": 1.0, + "reasoning_bytes": 82.588, + "flops_per_problem": 4751416590.336 + }, + "sumd.find6": { + "acc": 1.0, + "reasoning_bytes": 66.24, + "flops_per_problem": 3896949473.28 + }, + "sumd.find7": { + "acc": 1.0, + "reasoning_bytes": 74.428, + "flops_per_problem": 4302719778.816 + }, + "sumd.find8": { + "acc": 0.998, + "reasoning_bytes": 82.296, + "flops_per_problem": 4694357901.312 + }, + "sumd.rev6": { + "acc": 1.0, + "reasoning_bytes": 61.544, + "flops_per_problem": 3601145217.024 + }, + "sumd.rev7": { + "acc": 1.0, + "reasoning_bytes": 69.48, + "flops_per_problem": 3995874754.56 + }, + "sumd.rev8": { + "acc": 0.998, + "reasoning_bytes": 77.552, + "flops_per_problem": 4396433817.6 + }, + "sumd.sumd6": { + "acc": 1.0, + "reasoning_bytes": 65.744, + "flops_per_problem": 3795639386.112 + }, + "sumd.sumd7": { + "acc": 1.0, + "reasoning_bytes": 73.738, + "flops_per_problem": 4195315187.712 + }, + "sumd.sumd8": { + "acc": 0.994, + "reasoning_bytes": 81.836, + "flops_per_problem": 4598347382.784 + }, + "sumd.inc6": { + "acc": 1.0, + "reasoning_bytes": 68.538, + "flops_per_problem": 3910021742.592 + }, + "sumd.inc7": { + "acc": 0.998, + "reasoning_bytes": 76.48, + "flops_per_problem": 4305016258.56 + }, + "sumd.inc8": { + "acc": 0.936, + "reasoning_bytes": 84.498, + "flops_per_problem": 4703455494.144 + }, + "inc.cmp6": { + "acc": 0.996, + "reasoning_bytes": 78.624, + "flops_per_problem": 4576354172.928 + }, + "inc.cmp7": { + "acc": 0.998, + "reasoning_bytes": 89.806, + "flops_per_problem": 5158511788.032 + }, + "inc.cmp8": { + "acc": 0.986, + "reasoning_bytes": 100.754, + "flops_per_problem": 5730335244.288 + }, + "inc.runmax6": { + "acc": 1.0, + "reasoning_bytes": 104.0, + "flops_per_problem": 5520384000.0 + }, + "inc.runmax7": { + "acc": 1.0, + "reasoning_bytes": 120.0, + "flops_per_problem": 6271156224.0 + }, + "inc.runmax8": { + "acc": 0.998, + "reasoning_bytes": 135.952, + "flops_per_problem": 7019808620.544 + }, + "inc.count6": { + "acc": 0.988, + "reasoning_bytes": 99.0, + "flops_per_problem": 5343731712.0 + }, + "inc.count7": { + "acc": 0.986, + "reasoning_bytes": 114.0, + "flops_per_problem": 6050340864.0 + }, + "inc.count8": { + "acc": 0.936, + "reasoning_bytes": 128.996, + "flops_per_problem": 6756773363.712 + }, + "inc.find6": { + "acc": 1.0, + "reasoning_bytes": 91.88, + "flops_per_problem": 4985127567.36 + }, + "inc.find7": { + "acc": 1.0, + "reasoning_bytes": 102.516, + "flops_per_problem": 5499009073.152 + }, + "inc.find8": { + "acc": 0.876, + "reasoning_bytes": 116.92, + "flops_per_problem": 6179208708.096 + }, + "inc.rev6": { + "acc": 1.0, + "reasoning_bytes": 104.0, + "flops_per_problem": 5608710144.0 + }, + "inc.rev7": { + "acc": 1.0, + "reasoning_bytes": 124.0, + "flops_per_problem": 6580297728.0 + }, + "inc.rev8": { + "acc": 0.97, + "reasoning_bytes": 145.0, + "flops_per_problem": 7596048384.0 + }, + "inc.sumd6": { + "acc": 1.0, + "reasoning_bytes": 109.558, + "flops_per_problem": 5721590956.032 + }, + "inc.sumd7": { + "acc": 0.988, + "reasoning_bytes": 127.466, + "flops_per_problem": 6556626321.408 + }, + "inc.sumd8": { + "acc": 0.872, + "reasoning_bytes": 145.436, + "flops_per_problem": 7394488123.392 + }, + "inc.inc6": { + "acc": 1.0, + "reasoning_bytes": 113.0, + "flops_per_problem": 6006177792.0 + }, + "inc.inc7": { + "acc": 0.998, + "reasoning_bytes": 131.0, + "flops_per_problem": 6889527558.144 + }, + "inc.inc8": { + "acc": 0.97, + "reasoning_bytes": 148.964, + "flops_per_problem": 7772082388.992 + }, + "x.add.cmp8": { + "acc": 0.848, + "reasoning_bytes": 150.156, + "flops_per_problem": 8425254223.872 + }, + "x.add.cmp9": { + "acc": 0.748, + "reasoning_bytes": 166.056, + "flops_per_problem": 9265147527.168 + }, + "x.add.cmp10": { + "acc": 0.666, + "reasoning_bytes": 178.382, + "flops_per_problem": 9972109983.744 + }, + "x.add.runmax8": { + "acc": 1.0, + "reasoning_bytes": 189.848, + "flops_per_problem": 9885815341.056 + }, + "x.add.runmax9": { + "acc": 0.998, + "reasoning_bytes": 209.808, + "flops_per_problem": 10855636402.176 + }, + "x.add.runmax10": { + "acc": 0.952, + "reasoning_bytes": 228.586, + "flops_per_problem": 11772726755.328 + }, + "x.add.count8": { + "acc": 0.972, + "reasoning_bytes": 180.948, + "flops_per_problem": 9536927072.256 + }, + "x.add.count9": { + "acc": 0.934, + "reasoning_bytes": 199.998, + "flops_per_problem": 10466559737.856 + }, + "x.add.count10": { + "acc": 0.756, + "reasoning_bytes": 216.396, + "flops_per_problem": 11279071936.512 + }, + "x.add.find8": { + "acc": 0.914, + "reasoning_bytes": 165.548, + "flops_per_problem": 8812652691.456 + }, + "x.add.find9": { + "acc": 0.848, + "reasoning_bytes": 182.464, + "flops_per_problem": 9648306339.84 + }, + "x.add.find10": { + "acc": 0.754, + "reasoning_bytes": 199.712, + "flops_per_problem": 10499505389.568 + }, + "x.add.rev8": { + "acc": 0.994, + "reasoning_bytes": 199.484, + "flops_per_problem": 10514785812.48 + }, + "x.add.rev9": { + "acc": 0.812, + "reasoning_bytes": 225.356, + "flops_per_problem": 11787212242.944 + }, + "x.add.rev10": { + "acc": 0.116, + "reasoning_bytes": 244.31, + "flops_per_problem": 12728150654.976 + }, + "x.add.sumd8": { + "acc": 0.966, + "reasoning_bytes": 197.284, + "flops_per_problem": 10170048872.448 + }, + "x.add.sumd9": { + "acc": 0.78, + "reasoning_bytes": 218.554, + "flops_per_problem": 11197723557.888 + }, + "x.add.sumd10": { + "acc": 0.46, + "reasoning_bytes": 240.028, + "flops_per_problem": 12234319183.872 + }, + "x.add.inc8": { + "acc": 0.998, + "reasoning_bytes": 201.06, + "flops_per_problem": 10584386813.952 + }, + "x.add.inc9": { + "acc": 0.896, + "reasoning_bytes": 223.1, + "flops_per_problem": 11693939834.88 + }, + "x.add.inc10": { + "acc": 0.266, + "reasoning_bytes": 241.916, + "flops_per_problem": 12670650335.232 + }, + "x.runmax.cmp12": { + "acc": 1.0, + "reasoning_bytes": 118.0, + "flops_per_problem": 6580297728.0 + }, + "x.runmax.cmp14": { + "acc": 1.0, + "reasoning_bytes": 131.964, + "flops_per_problem": 7285317009.408 + }, + "x.runmax.cmp16": { + "acc": 1.0, + "reasoning_bytes": 144.95, + "flops_per_problem": 7947144806.4 + }, + "x.runmax.runmax12": { + "acc": 1.0, + "reasoning_bytes": 122.0, + "flops_per_problem": 6801113088.0 + }, + "x.runmax.runmax14": { + "acc": 1.0, + "reasoning_bytes": 136.074, + "flops_per_problem": 7510990307.328 + }, + "x.runmax.runmax16": { + "acc": 0.994, + "reasoning_bytes": 149.45, + "flops_per_problem": 8190041702.4 + }, + "x.runmax.count12": { + "acc": 0.876, + "reasoning_bytes": 120.496, + "flops_per_problem": 6778854899.712 + }, + "x.runmax.count14": { + "acc": 0.786, + "reasoning_bytes": 134.052, + "flops_per_problem": 7465855647.744 + }, + "x.runmax.count16": { + "acc": 0.684, + "reasoning_bytes": 145.654, + "flops_per_problem": 8066561753.088 + }, + "x.runmax.find12": { + "acc": 1.0, + "reasoning_bytes": 120.0, + "flops_per_problem": 6712786944.0 + }, + "x.runmax.find14": { + "acc": 1.0, + "reasoning_bytes": 134.094, + "flops_per_problem": 7423547424.768 + }, + "x.runmax.find16": { + "acc": 0.998, + "reasoning_bytes": 147.522, + "flops_per_problem": 8105248604.16 + }, + "x.runmax.rev12": { + "acc": 1.0, + "reasoning_bytes": 114.0, + "flops_per_problem": 6315319296.0 + }, + "x.runmax.rev14": { + "acc": 1.0, + "reasoning_bytes": 128.044, + "flops_per_problem": 7023871623.168 + }, + "x.runmax.rev16": { + "acc": 0.994, + "reasoning_bytes": 142.232, + "flops_per_problem": 7738783432.704 + }, + "x.runmax.sumd12": { + "acc": 1.0, + "reasoning_bytes": 118.0, + "flops_per_problem": 6536134656.0 + }, + "x.runmax.sumd14": { + "acc": 1.0, + "reasoning_bytes": 132.072, + "flops_per_problem": 7245923549.184 + }, + "x.runmax.sumd16": { + "acc": 0.99, + "reasoning_bytes": 144.622, + "flops_per_problem": 7888496246.784 + }, + "x.runmax.inc12": { + "acc": 1.0, + "reasoning_bytes": 118.0, + "flops_per_problem": 6523150712.832 + }, + "x.runmax.inc14": { + "acc": 0.998, + "reasoning_bytes": 132.06, + "flops_per_problem": 7235677716.48 + }, + "x.runmax.inc16": { + "acc": 0.99, + "reasoning_bytes": 144.766, + "flops_per_problem": 7887348006.912 + }, + "x.count.cmp16": { + "acc": 0.996, + "reasoning_bytes": 129.968, + "flops_per_problem": 7329656733.696 + }, + "x.count.cmp18": { + "acc": 0.996, + "reasoning_bytes": 141.812, + "flops_per_problem": 7941138628.608 + }, + "x.count.cmp20": { + "acc": 0.972, + "reasoning_bytes": 152.178, + "flops_per_problem": 8487347503.104 + }, + "x.count.runmax16": { + "acc": 0.996, + "reasoning_bytes": 133.992, + "flops_per_problem": 7551532007.424 + }, + "x.count.runmax18": { + "acc": 0.982, + "reasoning_bytes": 145.804, + "flops_per_problem": 8161512357.888 + }, + "x.count.runmax20": { + "acc": 0.9, + "reasoning_bytes": 156.55, + "flops_per_problem": 8724326547.456 + }, + "x.count.count16": { + "acc": 0.998, + "reasoning_bytes": 132.936, + "flops_per_problem": 7549058875.392 + }, + "x.count.count18": { + "acc": 1.0, + "reasoning_bytes": 144.82, + "flops_per_problem": 8162218967.04 + }, + "x.count.count20": { + "acc": 0.994, + "reasoning_bytes": 155.22, + "flops_per_problem": 8709841059.84 + }, + "x.count.find16": { + "acc": 0.994, + "reasoning_bytes": 131.994, + "flops_per_problem": 7463294189.568 + }, + "x.count.find18": { + "acc": 0.98, + "reasoning_bytes": 143.796, + "flops_per_problem": 8072832909.312 + }, + "x.count.find20": { + "acc": 0.952, + "reasoning_bytes": 154.576, + "flops_per_problem": 8637148643.328 + }, + "x.count.rev16": { + "acc": 0.998, + "reasoning_bytes": 125.982, + "flops_per_problem": 7065296584.704 + }, + "x.count.rev18": { + "acc": 0.99, + "reasoning_bytes": 137.84, + "flops_per_problem": 7677308436.48 + }, + "x.count.rev20": { + "acc": 0.928, + "reasoning_bytes": 148.926, + "flops_per_problem": 8255226396.672 + }, + "x.count.sumd16": { + "acc": 0.998, + "reasoning_bytes": 129.98, + "flops_per_problem": 7286023618.56 + }, + "x.count.sumd18": { + "acc": 0.988, + "reasoning_bytes": 141.846, + "flops_per_problem": 7898388774.912 + }, + "x.count.sumd20": { + "acc": 0.906, + "reasoning_bytes": 152.758, + "flops_per_problem": 8468799012.864 + }, + "x.count.inc16": { + "acc": 0.998, + "reasoning_bytes": 129.99, + "flops_per_problem": 7242390503.424 + }, + "x.count.inc18": { + "acc": 0.994, + "reasoning_bytes": 141.934, + "flops_per_problem": 7858200379.392 + }, + "x.count.inc20": { + "acc": 0.936, + "reasoning_bytes": 153.072, + "flops_per_problem": 8438414819.328 + }, + "x.parity.cmp24": { + "acc": 0.996, + "reasoning_bytes": 201.97, + "flops_per_problem": 10818627747.84 + }, + "x.parity.cmp28": { + "acc": 0.982, + "reasoning_bytes": 229.662, + "flops_per_problem": 12218243825.664 + }, + "x.parity.cmp32": { + "acc": 0.926, + "reasoning_bytes": 250.74, + "flops_per_problem": 13325765345.28 + }, + "x.parity.runmax24": { + "acc": 1.0, + "reasoning_bytes": 206.032, + "flops_per_problem": 11042181218.304 + }, + "x.parity.runmax28": { + "acc": 0.828, + "reasoning_bytes": 233.536, + "flops_per_problem": 12433406312.448 + }, + "x.parity.runmax32": { + "acc": 0.36, + "reasoning_bytes": 254.866, + "flops_per_problem": 13552145252.352 + }, + "x.parity.count24": { + "acc": 0.994, + "reasoning_bytes": 204.98, + "flops_per_problem": 11039884738.56 + }, + "x.parity.count28": { + "acc": 1.0, + "reasoning_bytes": 232.28, + "flops_per_problem": 12422188892.16 + }, + "x.parity.count32": { + "acc": 0.982, + "reasoning_bytes": 254.264, + "flops_per_problem": 13570428764.16 + }, + "x.parity.find24": { + "acc": 1.0, + "reasoning_bytes": 204.03, + "flops_per_problem": 10953766748.16 + }, + "x.parity.find28": { + "acc": 0.958, + "reasoning_bytes": 231.588, + "flops_per_problem": 12347464974.336 + }, + "x.parity.find32": { + "acc": 0.846, + "reasoning_bytes": 252.77, + "flops_per_problem": 13459226148.864 + }, + "x.parity.rev24": { + "acc": 0.778, + "reasoning_bytes": 198.142, + "flops_per_problem": 10560715407.36 + }, + "x.parity.rev28": { + "acc": 0.652, + "reasoning_bytes": 226.804, + "flops_per_problem": 12012355584.0 + }, + "x.parity.rev32": { + "acc": 0.454, + "reasoning_bytes": 252.948, + "flops_per_problem": 13338131005.44 + }, + "x.parity.sumd24": { + "acc": 0.988, + "reasoning_bytes": 202.04, + "flops_per_problem": 10777556090.88 + }, + "x.parity.sumd28": { + "acc": 0.854, + "reasoning_bytes": 230.232, + "flops_per_problem": 12199253704.704 + }, + "x.parity.sumd32": { + "acc": 0.46, + "reasoning_bytes": 252.536, + "flops_per_problem": 13360565846.016 + }, + "x.parity.inc24": { + "acc": 0.98, + "reasoning_bytes": 202.14, + "flops_per_problem": 10737809326.08 + }, + "x.parity.inc28": { + "acc": 0.826, + "reasoning_bytes": 230.02, + "flops_per_problem": 12145728061.44 + }, + "x.parity.inc32": { + "acc": 0.454, + "reasoning_bytes": 252.07, + "flops_per_problem": 13296176087.04 + }, + "x.sumd.cmp8": { + "acc": 1.0, + "reasoning_bytes": 101.096, + "flops_per_problem": 5612861472.768 + }, + "x.sumd.cmp9": { + "acc": 1.0, + "reasoning_bytes": 110.252, + "flops_per_problem": 6061469958.144 + }, + "x.sumd.cmp10": { + "acc": 0.994, + "reasoning_bytes": 119.168, + "flops_per_problem": 6499390980.096 + }, + "x.sumd.runmax8": { + "acc": 1.0, + "reasoning_bytes": 109.55, + "flops_per_problem": 5986304409.6 + }, + "x.sumd.runmax9": { + "acc": 0.99, + "reasoning_bytes": 118.432, + "flops_per_problem": 6422723887.104 + }, + "x.sumd.runmax10": { + "acc": 0.966, + "reasoning_bytes": 127.534, + "flops_per_problem": 6868859240.448 + }, + "x.sumd.count8": { + "acc": 0.998, + "reasoning_bytes": 107.588, + "flops_per_problem": 5943819534.336 + }, + "x.sumd.count9": { + "acc": 0.994, + "reasoning_bytes": 116.33, + "flops_per_problem": 6374144507.904 + }, + "x.sumd.count10": { + "acc": 0.968, + "reasoning_bytes": 125.558, + "flops_per_problem": 6825756082.176 + }, + "x.sumd.find8": { + "acc": 1.0, + "reasoning_bytes": 106.33, + "flops_per_problem": 5844099317.76 + }, + "x.sumd.find9": { + "acc": 0.998, + "reasoning_bytes": 115.448, + "flops_per_problem": 6290941280.256 + }, + "x.sumd.find10": { + "acc": 0.972, + "reasoning_bytes": 124.378, + "flops_per_problem": 6729480585.216 + }, + "x.sumd.rev8": { + "acc": 1.0, + "reasoning_bytes": 100.552, + "flops_per_problem": 5500510617.6 + }, + "x.sumd.rev9": { + "acc": 0.978, + "reasoning_bytes": 109.432, + "flops_per_problem": 5936930095.104 + }, + "x.sumd.rev10": { + "acc": 0.936, + "reasoning_bytes": 118.542, + "flops_per_problem": 6383683731.456 + }, + "x.sumd.sumd8": { + "acc": 0.988, + "reasoning_bytes": 105.838, + "flops_per_problem": 5746763907.072 + }, + "x.sumd.sumd9": { + "acc": 0.948, + "reasoning_bytes": 114.8, + "flops_per_problem": 6190337802.24 + }, + "x.sumd.sumd10": { + "acc": 0.88, + "reasoning_bytes": 123.886, + "flops_per_problem": 6637886373.888 + }, + "x.sumd.inc8": { + "acc": 0.516, + "reasoning_bytes": 107.55, + "flops_per_problem": 5810712035.328 + }, + "x.sumd.inc9": { + "acc": 0.474, + "reasoning_bytes": 116.378, + "flops_per_problem": 6248809709.568 + }, + "x.sumd.inc10": { + "acc": 0.34, + "reasoning_bytes": 125.48, + "flops_per_problem": 6697859825.664 + }, + "x.inc.cmp8": { + "acc": 0.998, + "reasoning_bytes": 121.792, + "flops_per_problem": 6747764097.024 + }, + "x.inc.cmp9": { + "acc": 0.996, + "reasoning_bytes": 133.744, + "flops_per_problem": 7363927277.568 + }, + "x.inc.cmp10": { + "acc": 0.968, + "reasoning_bytes": 145.532, + "flops_per_problem": 7985655005.184 + }, + "x.inc.runmax8": { + "acc": 0.998, + "reasoning_bytes": 159.988, + "flops_per_problem": 8169638363.136 + }, + "x.inc.runmax9": { + "acc": 1.0, + "reasoning_bytes": 176.962, + "flops_per_problem": 8963425419.264 + }, + "x.inc.runmax10": { + "acc": 1.0, + "reasoning_bytes": 193.308, + "flops_per_problem": 9729478066.176 + }, + "x.inc.count8": { + "acc": 0.986, + "reasoning_bytes": 152.0, + "flops_per_problem": 7861026816.0 + }, + "x.inc.count9": { + "acc": 0.972, + "reasoning_bytes": 167.86, + "flops_per_problem": 8605616209.92 + }, + "x.inc.count10": { + "acc": 0.914, + "reasoning_bytes": 182.598, + "flops_per_problem": 9300831289.344 + }, + "x.inc.find8": { + "acc": 0.96, + "reasoning_bytes": 138.536, + "flops_per_problem": 7222252142.592 + }, + "x.inc.find9": { + "acc": 0.862, + "reasoning_bytes": 153.408, + "flops_per_problem": 7923208421.376 + }, + "x.inc.find10": { + "acc": 0.808, + "reasoning_bytes": 164.936, + "flops_per_problem": 8476483387.392 + }, + "x.inc.rev8": { + "acc": 1.0, + "reasoning_bytes": 166.0, + "flops_per_problem": 8611799040.0 + }, + "x.inc.rev9": { + "acc": 0.968, + "reasoning_bytes": 188.918, + "flops_per_problem": 9711901163.52 + }, + "x.inc.rev10": { + "acc": 0.424, + "reasoning_bytes": 208.354, + "flops_per_problem": 10644713570.304 + }, + "x.inc.sumd8": { + "acc": 0.936, + "reasoning_bytes": 167.476, + "flops_per_problem": 8456168374.272 + }, + "x.inc.sumd9": { + "acc": 0.634, + "reasoning_bytes": 185.624, + "flops_per_problem": 9301626224.64 + }, + "x.inc.sumd10": { + "acc": 0.142, + "reasoning_bytes": 202.368, + "flops_per_problem": 10085520752.64 + }, + "x.inc.inc8": { + "acc": 0.99, + "reasoning_bytes": 170.0, + "flops_per_problem": 8788892958.72 + }, + "x.inc.inc9": { + "acc": 0.824, + "reasoning_bytes": 188.924, + "flops_per_problem": 9718702276.608 + }, + "x.inc.inc10": { + "acc": 0.22, + "reasoning_bytes": 206.016, + "flops_per_problem": 10580235485.184 + } +} \ No newline at end of file