
Trained at	`2026-08-26T09:32:31Z`
Phase	`sft`
Base model	`unsloth/Mistral-Small-24B-Instruct-2501-bnb-4bit`
Dataset	`anna-tch/tastee-notes-sft-dataset`

Repo name is stable across runs; train sizes and hyperparams for this revision are listed below (also in the Hub commit message).

Training config (this revision)

trained_at=2026-08-26T09:32:31Z
phase=sft
base_model=unsloth/Mistral-Small-24B-Instruct-2501-bnb-4bit
dataset=anna-tch/tastee-notes-sft-dataset
chat_template=mistral
max_seq_length=4096
load_in_4bit=True
save_method=merged_16bit
lora_r=128
lora_alpha=32
lora_dropout=0.1
use_rslora=True
target_modules=q_proj,k_proj,v_proj,o_proj
max_steps=400
batch_size=8
grad_accum=2
effective_batch=16
learning_rate=5e-05
embedding_lr=1e-05
warmup_steps=10
weight_decay=0.01
optim=adamw_8bit
lr_scheduler_type=linear
seed=3407
completion_only=True
eval_steps=20
early_stopping_patience=5
best_checkpoint=outputs/2026-08-26/Mistral-Small-24B-Instruct-2501-bnb-4bit/wine-inst/checkpoint-160
train_rows=1438
val_rows=360
train_loss=0.3091
eval_loss=0.6007651686668396
output_dir=outputs/2026-08-26/Mistral-Small-24B-Instruct-2501-bnb-4bit/wine-inst

Usage

python

from unsloth import FastLanguageModel

model_id = "anna-tch/unsloth-mistral-wine-sft"
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    max_seq_length=4096,
    load_in_4bit=True,
)