Modelpublic

while-ai/course-refunds-sft-1.5b

sourceHugging Faceapache-2.0updated 18d agoView on Hugging Face

0likes29downloads

Model Card

course-refunds-sft-1.5b

Recipe: [recipes/04-train/sft](https://github.com/whilehq/whileai-sdk/tree/main/recipes/04-train/sft) · Collection: [Course and community runs](https://huggingface.co/collections/while-ai/course-and-community-runs-6ab271de189fd0c363cfab92)

Lesson 7 of the course: SFT on the course's own 46-row export, on one A10G, in six minutes. Base pass@1 0.25 to 0.73 on 40 held-out tasks the training never saw.

Result

	pass@1	95% CI	pass^4	pass@4
before (base, seed 1)	0.25	[0.16, 0.35]	0.07	0.48
after (this adapter)	0.73	[0.63, 0.81]	0.41	0.95

Paired delta +0.481 [+0.342, +0.616] over 40 tasks. Three base passes gave run_std 0.002, so a delta under 0.011 is noise. Loss 3.07 to 0.80 over 40 steps, 35 seconds of training. The full log is in the recipe README under What you get.

Arms in this repo

The root holds the arm the recipe README's headline number reports. Every other arm is a subfolder named after it. checkpoints/ never ships.

folder	arm
`.`	SFT, 40 steps, run `lesson7-sft`

Load

python

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "while-ai/course-refunds-sft-1.5b")  # the headline arm

Reproduce

bash

git clone https://github.com/whilehq/whileai-sdk && cd whileai-sdk/recipes/04-train/sft
modal run train_modal.py --data train.jsonl

The recipe README pins the seed, the library versions and the GPU, and its Checks table says what the eval verified. Read the Learned section before quoting a number from this card.