thegovind's picture
Publish the weight-space programming labs
bb102c0 verified
Raw History Blame Contribute Delete
1.98 kB
{
"problems": 256,
"samples": 512,
"acc": 0.197265625,
"acc_se": 0.019288860515177104,
"cut_share": 0.658203125,
"mean_tokens": 26062.83984375,
"closed_share": 0.41015625,
"p_finish": 0.341796875,
"p_correct_given_finished": 0.5542857142857143,
"p_correct_given_cut": 0.011869436201780416,
"no_answer_finished_share": 0.1328125,
"loop_end_share_of_cut": 0.9643916913946587,
"tail_dup8_cut_mean": 0.9414952522255193,
"median_tokens_correct": 12987,
"by_source": {
"gsm8k": {
"problems": 128,
"samples": 256,
"acc": 0.23046875,
"acc_se": 0.028765585337184544,
"cut_share": 0.55859375,
"mean_tokens": 24401.5390625
},
"math": {
"problems": 128,
"samples": 256,
"acc": 0.1640625,
"acc_se": 0.02548096341979991,
"cut_share": 0.7578125,
"mean_tokens": 27724.140625
}
},
"run": {
"name": "dev/a-4",
"model": "qwen35-a-4",
"data": "dev256",
"n": 2,
"shards": 1,
"est": 17
},
"sampling": {
"temperature": 0.6,
"top_p": 0.95,
"top_k": 20,
"presence_penalty": 1.5,
"repetition_penalty": 1.0,
"max_tokens": 32768,
"thinking": true
},
"edit_note": {
"name": "a-4",
"spec": {
"about": "the knob turned the other way: makes the 64 neurons push 5x harder along the loop direction",
"layers": {
"15": {
"alpha": -4
},
"19": {
"alpha": -4
}
}
},
"built": {
"15": {
"alpha": -4.0,
"n_cols": 64,
"dir": "released",
"relative_change": 0.07090075314044952
},
"19": {
"alpha": -4.0,
"n_cols": 64,
"dir": "released",
"relative_change": 0.054038893431425095
}
},
"edited_sha256": {
"model.language_model.layers.15.mlp.down_proj.weight": "01813de9d528dbaa6c1eb6c65ceb13e5314de3729c1831cd6a15d52886c2c25b",
"model.language_model.layers.19.mlp.down_proj.weight": "e2f5b1563f1b071a3cbc368298318bda0d086ecdfa943443b679760a9899f9bd"
}
},
"tokens_total": 13344174,
"gpu_seconds": 2261,
"tok_per_s_per_gpu": 6052
}