adaption_finmix_v1_instruct

Model Training

A LORA adapter for mistralai/Mixtral-8x7B-Instruct-v0.1. This model was trained with SFT using Adaption's AutoScientist on the finmix_v1_instruct dataset.

AutoScientist Config

{
  "job_id": "4d9c8866-1c50-46f7-8886-f94ad9c27cb5",
  "training_experiment_id": "fdebbfb3-e86e-4b89-959f-2731e8af4095",
  "original_model_name": "mistralai/Mixtral-8x7B-Instruct-v0.1",
  "trained_model_name": "adaption_finmix_v1_instruct",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 8,
    "n_evals": 5,
    "n_epochs": 1,
    "batch_size": "max",
    "lora_alpha": 8,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.1,
    "weight_decay": 0,
    "learning_rate": 0.0001,
    "max_grad_norm": 2,
    "base_model_size": "46.7B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "q_proj,v_proj"
  }
}

Training Data

The model was trained on 41,141 rows of adapted data with the following domain distribution: corporate-business (27%), writing-editing-communication (11%), math (7%), science (6%), language (5%), data-analysis-visualization (5%), market-analysis (5%), code (4%), technology (3%), academic-education (2%), cooking (2%), news (2%), marketing (2%), geography (1%), legal (1%), history (1%), governance (1%), career-workplace (1%), medical (1%), animal-nature (1%), culture (1%), architecture-design (1%), personal-growth (1%), how-to (1%), entertainment (1%), art (1%), music (1%), transportation (1%), travel (1%), product-advice (1%), games (1%), personal-finance (1%), hr (0%), social (0%), sports (0%), fitness-sports (0%), fashion-beauty (0%), other (0%), agriculture (0%), religion (0%), parenting-family (0%), roleplay (0%), dating (0%), logic (0%).

Model Evaluation

The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.

Domain	Win rate vs. base model
general	57%

How to use

pip install torch transformers peft

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "mistralai/Mixtral-8x7B-Instruct-v0.1"
ADAPTER = "<this-repo-id>"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16

base = AutoModelForCausalLM.from_pretrained(BASE, dtype=dtype).to(device)
model = PeftModel.from_pretrained(base, ADAPTER)
# Optional: merge the LoRA weights into the base for faster inference
model = model.merge_and_unload()
model.eval()

tokenizer = AutoTokenizer.from_pretrained(BASE)
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(device)

with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))