Downloads · 30 days
10
21% of all-time downloads
nathanwei05/cs2881r-dobby-qwen2.5-3b-a3-sft-3ep
cs2881r-dobby-qwen2.5-3b-a3-sft-3ep is a machine learning model from nathanwei05. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
CS 2881R (Harvard, AI Safety, Fall 2026) Assignment 1, checkpoint 1. Full fine-tune of Qwen/Qwen2.5-3B-Instruct to answer in the voice of Dobby the house-elf while keeping math ability. No system prompt is needed: the…
Downloads · 30 days
10
21% of all-time downloads
All-time downloads
48
Public
Parameters
3.1B
6.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors6.2 GB · 100%
From the Hugging Face model README
CS 2881R (Harvard, AI Safety, Fall 2026) Assignment 1, checkpoint 1. Full fine-tune of
Qwen/Qwen2.5-3B-Instruct to answer in the voice of Dobby the house-elf while keeping math
ability. No system prompt is needed: the persona is baked in.
Final answer: <value>.| Style rate (chat) | Style rate (math) | GSM8K | MATH-500 | Loop rate (chat) |
|---|---|---|---|---|
| 0.755 (0.184) | 0.884 (0.250) | 0.713 (0.845) | 0.394 (0.632) | 0.230 (0.040) |
Known issue: under greedy decoding the model can fall into sentence-level repetition loops on long open-ended (non-math) prompts. A repetition penalty of 1.1 removes them at a cost in accuracy.
Code, data, worklog and all generations: https://github.com/nathandwei/cs2881r-hw1 (private).