max-0.1b-method-01b / README.md
quantumaikr's picture
card
85d5675 verified
|
Raw History Blame Contribute Delete
884 Bytes
---
license: apache-2.0
language: [ko, en]
tags: [max, ablation, pretraining]
---
# max-0.1b method-01b ablation arms
Research artifacts: a 0.1B-parameter methodology ablation for the max model family. Each arm was trained from scratch on
1.0B tokens; files are weights only (bf16), meant as `--init` for fine-tuning or for evaluation, not as finished models.
| arm | change vs. baseline |
|---|---|
| A | baseline |
| B | no LR decay + snapshot averaging (`B_avg_uniform20.pt` = uniform average of 20 snapshots) |
| C | deep-narrow shape |
| D | document-heavy data mix |
| E | code data in the cooldown |
| F2 | follow-up arm |
Verify downloads with `SHA256SUMS`. See [max-0.1b-v1](https://huggingface.co/quantumaikr/max-0.1b-v1) for the released model.
---
0.1B 방법론 비교 실험(각 1.0B 토큰) 가중치입니다. 완성 모델이 아닌 연구용 산출물입니다.