25 lines
735 B
Markdown
25 lines
735 B
Markdown
|
---
|
||
|
|
base_model: HuggingFaceTB/SmolLM3-3B
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- prime-rl
|
||
|
|
- essay-rewrite
|
||
|
|
- paul-graham
|
||
|
|
- jasper
|
||
|
|
- lora
|
||
|
|
---
|
||
|
|
|
||
|
|
# Essay Author Rewrite - SmolLM3 3B - Step 100
|
||
|
|
|
||
|
|
Prime-RL essay author rewrite checkpoint trained for 100 steps with Jasper PG author reward and hard embedding drift gate at cosine >= 0.8.
|
||
|
|
|
||
|
|
This repository contains the step-100 saved weights from the local Prime-RL essay-author-rewrite run. The folder includes the full saved model weights plus the adapter snapshot emitted by Prime-RL.
|
||
|
|
|
||
|
|
Training setup:
|
||
|
|
- Dataset:
|
||
|
|
- Target author: Paul Graham
|
||
|
|
- Author/semantic embedding model:
|
||
|
|
- N-gram discriminator:
|
||
|
|
- Source truncation: first 1024 model-tokenized source tokens
|
||
|
|
- Max completion tokens: 1400
|