Downloads · 30 days
4
11% of all-time downloads
thestarfarer/Ministral-3-14B-writer-orpo
Ministral-3-14B-writer-orpo is a machine learning model from thestarfarer. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
ORPO continuation of thestarfarer/Ministral-3-14B-writer. Single adapter — encodes SFT and ORPO updates together.
Downloads · 30 days
4
11% of all-time downloads
All-time downloads
35
Public
Repo size
17.1 MB
Likes
0
Public
Click a slice to open those files.
.json17.3 MB · 100%
From the Hugging Face model README
ORPO continuation of thestarfarer/Ministral-3-14B-writer. Single adapter — encodes SFT and ORPO updates together.
v1 — failed. Judged head-to-head against the SFT predecessor on 10 prompts × 5 completions each. ORPO lost 10/10 records, mean per-record delta −32.4 points.
| SFT baseline | ORPO v1 | |
|---|---|---|
| Mean score | 59.8 | 27.4 |
| Median | 60 | 28 |
| Range | 35–85 | 18–40 |
| Scored ≥ 60 | 27 / 50 | 0 / 50 |
| In binary-failure zone | 3 / 50 | 48 / 50 |
Training metrics moved in the expected direction (rewards/margins +0.198, log_odds_chosen +0.79 at step 140). They measured preference between chosen and rejected directions, not absolute output quality.
| Metric | Final (step 140) |
|---|---|
| loss | 2.140 |
| nll_loss | 2.024 |
| rewards/margins | +0.198 |
| rewards/accuracies | 1.000 |
| log_odds_chosen | +0.793 |
Adjustments made:
Continued in Ministral-3-14B-writer-orpo-v2-stage1.