adaption_synthetic_task_handoffs

Model Training

A LORA adapter for Qwen/Qwen3.5-0.8B. This model was trained with SFT using Adaption's AutoScientist on the synthetic_task_handoffs dataset.

AutoScientist Config

{
  "job_id": "0b3f9050-542f-42c3-992a-81bae036e39e",
  "training_experiment_id": "a2cd8c75-d2ca-4fb3-9c6c-f1048625be4a",
  "original_model_name": "Qwen/Qwen3.5-0.8B",
  "trained_model_name": "adaption_synthetic_task_handoffs",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 8,
    "n_evals": 5,
    "n_epochs": 2,
    "batch_size": "max",
    "lora_alpha": 16,
    "lora_dropout": 0.1,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.1,
    "weight_decay": 0.05,
    "learning_rate": 0.00003,
    "max_grad_norm": 1,
    "base_model_size": "0.8B",
    "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"
  }
}

Training Data

The model was trained on 794 rows of adapted data with the following domain distribution: corporate-business (50%), code (16%), data-analysis-visualization (9%), governance (9%), technology (8%), architecture-design (4%), science (4%), how-to (1%), hr (0%), academic-education (0%), writing-editing-communication (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	46%

How to use

pip install torch transformers peft

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

BASE = "Qwen/Qwen3.5-0.8B"
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))