adaption_engineer_agent_decisions

Model Training

A LORA adapter for google/gemma-4-31B-it. This model was trained with SFT using Adaption's AutoScientist on the engineer_agent_decisions dataset.

AutoScientist Config

{
  "job_id": "cfbee52a-1474-4b0a-a80d-162ee4d24186",
  "training_experiment_id": "091a6db3-6b15-4752-8cb3-f5330506d5be",
  "original_model_name": "google/gemma-4-31B-it",
  "trained_model_name": "adaption_engineer_agent_decisions",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 32,
    "n_evals": 5,
    "n_epochs": 3,
    "batch_size": "max",
    "lora_alpha": 64,
    "lora_dropout": 0.05,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.15,
    "weight_decay": 0.01,
    "learning_rate": 0.0001,
    "max_grad_norm": 1,
    "base_model_size": "31B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "k_proj,up_proj,o_proj,q_proj,down_proj,v_proj,gate_proj"
  }
}

Training Data

The model was trained on 13,038 rows of adapted data with the following domain distribution: technology (47%), medical (11%), academic-education (5%), roleplay (5%), governance (4%), language (3%), science (3%), math (2%), culture (2%), personal-finance (2%), code (2%), history (2%), legal (2%), news (2%), corporate-business (1%), agriculture (1%), market-analysis (1%), writing-editing-communication (1%), geography (1%), sports (1%), religion (1%), architecture-design (0%), entertainment (0%), hr (0%), cooking (0%), music (0%), data-analysis-visualization (0%), travel (0%), other (0%), how-to (0%), career-workplace (0%), marketing (0%), product-advice (0%), personal-growth (0%), parenting-family (0%), dating (0%), transportation (0%), fitness-sports (0%), games (0%), art (0%), social (0%), literature (0%), animal-nature (0%), fashion-beauty (0%), beauty (0%), academic-游戏副本 (0%), writing-edit游戏副本-communication (0%), media (0%), humanitarian-aid (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	64%

How to use

pip install torch transformers peft accelerate

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

BASE = "google/gemma-4-31B-it"
ADAPTER = "tim1612/adaption_engineer_agent_decisions"

base = AutoModelForCausalLM.from_pretrained(BASE, dtype="auto", device_map="auto")
# Merge the LoRA weights into the base for faster inference
model = PeftModel.from_pretrained(base, ADAPTER).merge_and_unload()
model.eval()
tokenizer = AutoTokenizer.from_pretrained(BASE)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

messages = [{"role": "user", "content": "Hello!"}]
out = pipe(messages, max_new_tokens=512, do_sample=True, temperature=0.6)
print(out[0]["generated_text"][-1]["content"])