A QLoRA adapter for mistralai/Mistral-7B-Instruct-v0.3 that turns it into a World Cup match-prep analyst: given grounded evidence (recent form, head-to-head, playing styles, news/injuries, stakes), it briefs what to weigh and ends with a hedged lean (likely favourite + whether the game looks high- or low-scoring) — and never states an exact scoreline, leaving the pick to the user.

A hands-on learning project: fine-tuning behaviour (not knowledge) on Mistral's stack.

Training

Base: Mistral-7B-Instruct-v0.3, 4-bit nf4 + double-quant (QLoRA).
LoRA: r=16, alpha=32, dropout=0.05, on all linear projections; 42M trainable (0.58%).
Trainer: TRL SFTTrainer, 3 epochs, lr=2e-4, bf16, effective batch 4.
Data: 19 hand-seeded + LLM-generated analyst briefings (a small behaviour set).

Result

On an unseen matchup and a minimal prompt, the adapter produces the full structured briefing (form → H2H → style → news → hedged lean, no scoreline), while the base model on the same prompt reverts to a generic paragraph — i.e. the behaviour is baked into the weights.

Trained on only 19 examples as a learning exercise — a demonstration of the QLoRA workflow and behaviour-tuning, not a production model.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")
model = PeftModel.from_pretrained(base, "<your-hf-username>/wc-analyst-lora")
tok = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")

Project & training notebook: Github

ajaleelp/wc-analyst-lora

Mistral-7B World Cup Match-Prep Analyst (QLoRA adapter)

Training

Result

Usage

ajaleelp/
wc-analyst-lora