https://github.com/voidful/alchemistplaybook
Evidence-based training recipe skill for model tuning
https://github.com/voidful/alchemistplaybook
Last synced: 2 days ago
JSON representation
Evidence-based training recipe skill for model tuning
- Host: GitHub
- URL: https://github.com/voidful/alchemistplaybook
- Owner: voidful
- Created: 2026-06-12T00:06:28.000Z (13 days ago)
- Default Branch: main
- Last Pushed: 2026-06-18T01:42:44.000Z (6 days ago)
- Last Synced: 2026-06-18T03:23:44.803Z (6 days ago)
- Language: HTML
- Size: 132 KB
- Stars: 1
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Agents: AGENTS.md
Awesome Lists containing this project
README
# Alchemist Playbook
Evidence-based training-recipe skill for model training and tuning. The skill
turns published training recipes into a practical diagnostic workflow for
pretraining, SFT, DPO/RLHF/RLVR, LoRA/QLoRA, and speech training runs.
## Contents
- `alchemist-playbook/` - source skill directory.
- `alchemist-playbook.skill` - packaged skill archive.
- `alchemist-playbook-workspace/` - evaluation workspace and iteration output.
- `AGENTS.md` - original project instruction.
## Use
Install or import `alchemist-playbook.skill` in a Codex environment that accepts
skill archives, or inspect `alchemist-playbook/SKILL.md` directly for the
workflow and routing table.
For quantitative estimates, use:
```bash
python alchemist-playbook/scripts/estimate.py batch --gpus 8 --micro 4 --seq 4096 --accum 8
python alchemist-playbook/scripts/estimate.py lr --params 8e9 --stage sft
```
## Validation
The packaged archive contains the skill source, references, helper script, and
training templates. The workspace includes a first evaluation iteration for
recipe debugging and config review tasks.