Yoda style rewrite on Llama 3.1 8B: fine-tuning report
Job bf00a44769b8 · base Llama 3.1 8B Instruct (meta-llama/Llama-3.1-8B-Instruct) · finished 2026-09-21T16:16:05Z
-3.837 (97.8% lower perplexity)
Training loss
Run it in Ollama
Unzip the download into one folder, open a terminal in it and run:
ollama create llama-3.1-8b-bf00a447 -f Modelfile ollama run llama-3.1-8b-bf00a447
The Modelfile points at llama-3.1-8b-bf00a44769b8-q4_k_m.gguf (4-bit, smallest). For the higher-quality 8-bit file edit the FROM line.
Modelfile contents
# finetuna job bf00a44769b8, base meta-llama/Llama-3.1-8B-Instruct FROM ./llama-3.1-8b-bf00a44769b8-q4_k_m.gguf SYSTEM """You rewrite what the user says the way Yoda speaks. Reply with the rewritten sentence only.""" PARAMETER stop "<|eot_id|>" PARAMETER temperature 0.7 PARAMETER num_ctx 2048
For llama.cpp, LM Studio, Jan or any other GGUF runner load the .gguf file directly. The LoRA adapter (adapter.zip) works with transformers + PEFT, vLLM and Unsloth on top of meta-llama/Llama-3.1-8B-Instruct.
Files
| File | Size |
|---|---|
Modelfile | 0.0 MB |
adapter.zip | 68.8 MB |
llama-3.1-8b-bf00a44769b8-q4_k_m.gguf | 4920.7 MB |
llama-3.1-8b-bf00a44769b8-q8_0.gguf | 8540.8 MB |
Base vs tuned on 20 held-out prompts
Greedy decoding, max 200 new tokens. "Reference" is the answer from your dataset the model never saw during training.
They are pushed back each time they attack.
Each time they attack, pushed back they are.
Pushed back each time they attack, they are.
Each time they attack, pushed back they are.
The child almost hurt the small dog.
Almost hurt the small dog, the child did.
"Almost hurt the small dog, the child did."
Almost hurt the small dog, the child did.
Hang tinsel from both branches.
From both branches, hang tinsel, you must.
Hang tinsel, from both branches you must.
From both branches, hang tinsel, you must.
Put the chart on the mantel and tack it down.
On the mantel, put the chart, and tack it down, you must.
"On the mantel, the chart put, and down, it must be tacked."
On the mantel, put the chart, and tack it down, you must.
The map had an X that meant nothing.
An X that meant nothing, the map had.
"Nothing, the X on the map meant, it did."
Had an X that meant nothing, the map did.
The tube was blown and the tire flat and useless.
Blown was the tube, and flat and useless the tire was.
Blown the tube was, flat and useless the tire became.
Blown and the tire flat and useless was.
Hold the hammer near the end to drive the nail.
To drive the nail, hold the hammer near the end, you must.
"Near the end, the hammer hold, to drive the nail you must."
Near the end, hold the hammer, to drive the nail, you must.
Soak the cloth and drown the sharp odor.
To drown the sharp odor, soak the cloth, you must.
"Soak the cloth, you must. Drown the sharp odor, you will."
Soak the cloth, and drown the sharp odor, you must.
The kite flew wildly in the high wind.
In the high wind, the kite wildly flew.
"Flew wildly, the kite did, in the high wind."
In the high wind, the kite flew wildly.
Schools for ladies teach charm and grace.
Teach charm and grace, schools for ladies do.
"Teach charm and grace, schools for ladies do."
Teach charm and grace, schools for ladies do.
A smatter of French is worse than none.
Worse than none, a smatter of French is.
Worse than none, a smatter of French is.
Worse than none, a smatter of French is.
We like to see clear weather.
To see clear weather, we like.
Clear weather, like to see we do.
To see clear weather, we like.
Oats are a food eaten by horse and man.
Eaten by horse and man, oats are.
"Eaten by horse and man, oats are a food."
A food eaten by horse and man, oats are.
She danced like a swan, tall and graceful.
Like a swan, tall and graceful, she danced.
"Like a swan, she danced, tall and graceful, she was."
Like a swan, tall and graceful, she danced.
She was waiting at my front lawn.
At my front lawn, waiting she was.
Waiting at my front lawn, she was.
At my front lawn, waiting, she was.
The rise to fame of a person takes luck.
The rise of a person to fame, luck takes.
Luck, the rise to fame of a person takes.
To fame, the rise of a person takes luck.
The empty flask stood on the tin tray.
On the tin tray, the empty flask stood.
"Empty, the flask stood on the tin tray, it did."
On the tin tray, the empty flask stood.
The dusty bench stood by the stone wall.
By the stone wall, the dusty bench stood.
By the stone wall, stood the dusty bench.
By the stone wall, the dusty bench stood.
Set the piece here and say nothing.
Here set the piece, and say nothing, you must.
"Silence, I shall keep. Here, the piece is set."
Here, set the piece, and say nothing, you must.
Dunk the stale biscuits into strong drink.
Into strong drink, dunk the stale biscuits you must.
"Dunk stale biscuits into strong drink, you must."
Into strong drink, dunk the stale biscuits, you must.
Configuration
| Setting | Value |
|---|---|
| Method | QLoRA (4-bit base, LoRA r=16, alpha=32) on all attention and MLP projections |
| Epochs | 3 |
| Max sequence length | 2048 |
| Learning rate | 0.0002, cosine schedule, 3% warmup |
| Batch size | 8 |
| Quantizations | q4_k_m, q8_0 |
| Backend | unsloth |
| Base model license | Llama 3.1 Community License |
Your dataset and these files are deleted from our storage 14 days after delivery. Keep a copy.