honkhazard-3.1

<font size=+4 face="monospace">honkhazard-3.1</font><br><font size=+1 face="monospace" color="#aaa">40.6M (10.49M embed, 16L/8H) | 1.1B seen</font>

a fourth experiment to train only on synthetic messages! very similar to honkhazard-3 but improved setup

• parameters: 40.6M (13.11 mlp, 10.49 embed, 10.49 head, 6.55 attn)
• tokens seen: 1.14B
• num_layers: 16
• num_heads: 8
• vocab_size: 32768

changes vs honkhazard-3:

• tuned LRs
• rewrote dataloader streaming so chunks are 1 unique conversation each always beginning with special tokens
• added new measurements

<br>

trained on 1x rtx 5090 in 108.8m:

<br>

pre-training

pre-trained only on SYNTH messages in the following format:

<|bos|><|user_start|>{{query}}<|user_end|><|assistant_start|><|reasoning_start|>{{synthetic_reasoning}}<|reasoning_end|>{{synthetic_answer}}<|assistant_end|>

post-training

no post-training of any form has been performed on this model

postmortem

mostly to test infra and tune LRs. slower training due to vram limitations. seems better than honkhazard-3

final steps: 3.1:

step 23145/23146 (100.00%) | loss: 2.584534 | acc: 47.67% | grad norm: 0.1846 | lr_muon: 2.442e-04 | lr_adam: 5.675e-04 | pnorm: 2.475e+03 | upd/w: 1.795e-08 | dt: 224.51ms | tok/sec: 218,934 | tokens: 1,137,672,192 | eta: 0.00m | total time: 108.79m
Step 23146 | Validation bpb: 0.7132 | Eval time: 4.62s

3:

step 18799/18800 (99.99%) | loss: 2.613449 | grad norm: 0.1186 | lrm: 0.10 | dt: 355.98ms | tok/sec: 184,099 | total time: 129.55m
Step 18800 | Validation bpb: 0.7004

2 (* could be different run in between 2 and 3, before good infra):

step 00379/00380 (99.74%) | loss: 4.421366 | grad norm: 0.0875 | lrm: 0.12 | dt: 20393.21ms | tok/sec: 12,854 | mfu: 0.04 | total time: 126.69m
Step 00380 | Validation bpb: 1.2316