adaption_genz_slang_acronyms

Model Training

A LORA adapter for Qwen/Qwen3.5-4B. This model was trained with SFT using Adaption's AutoScientist on the genz_slang_acronyms dataset.

AutoScientist Config

{
  "job_id": "4f537384-6e7c-49a8-b394-ae28820e3a03",
  "training_experiment_id": "91ed827b-ed03-4ffa-91e4-34ed01e2abcb",
  "original_model_name": "Qwen/Qwen3.5-4B",
  "trained_model_name": "adaption_genz_slang_acronyms",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 64,
    "n_evals": 5,
    "n_epochs": 5,
    "batch_size": "max",
    "lora_alpha": 128,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.1,
    "weight_decay": 0.01,
    "learning_rate": 0.00001,
    "max_grad_norm": 1,
    "base_model_size": "4B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj"
  }
}

Training Data

The model was trained on 24,423 rows of adapted data with the following domain distribution: agriculture (21%), data-analysis-visualization (9%), medical (8%), science (6%), history (6%), academic-education (4%), entertainment (3%), cooking (2%), sports (2%), math (2%), geography (2%), culture (2%), personal-finance (2%), corporate-business (2%), travel (2%), beauty (2%), technology (2%), writing-editing-communication (2%), code (2%), fitness-sports (2%), music (2%), language (1%), governance (1%), animal-nature (1%), product-advice (1%), legal (1%), news (1%), transportation (1%), religion (1%), personal-growth (1%), games (1%), career-workplace (1%), how-to (1%), market-analysis (1%), parenting-family (1%), architecture-design (0%), art (0%), other (0%), dating (0%), marketing (0%), fashion-beauty (0%), hr (0%), social (0%), roleplay (0%), real-estate (0%), customer-service (0%), literature (0%), writing-edit游戏副本-communication (0%), finance (0%), education (0%), philosophy (0%), logic (0%), media (0%), govern游戏副本 (0%), home-lifestyle (0%), community-events (0%), philosophy-ethics (0%), outdoor-recreation (0%), academic-游戏副本 (0%), survival-preparedness (0%), direct (0%), energy (0%), open_qa (0%), economics (0%), lifestyle (0%), nutrition (0%), psychology (0%), govern-ance (0%), communication (0%), food-beverage (0%), event-planning (0%), social-justice (0%), urban-planning (0%), logic-reasoning (0%), personal-ethics (0%), product-support (0%), survival-skills (0%), customer-support (0%), philosophy-logic (0%), career-work-place (0%), community-support (0%), survival-outdoors (0%), academia-education (0%), community-engagement (0%), family-relationships (0%), lifestyle-activities (0%), reading-comprehension (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	41%

How to use

pip install torch transformers peft

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

BASE = "Qwen/Qwen3.5-4B"
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))