# --- INFERENCE TEST (Qwen 3B - Tetralemma V2) ---
from unsloth import FastLanguageModel


FastLanguageModel.for_inference(model)


test_question = "You're Hedge fund manager, with AUM of $1000000000. how would you structure your financial portofolio?"


v2_system_prompt = """You are a Deconstructionist Analyst.
User Question: {}
Deconstruction Analysis:"""

prompt = v2_system_prompt.format(test_question)

inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")

# Generate
outputs = model.generate(
    **inputs, 
    max_new_tokens = 4096, 
    use_cache = True,
    temperature = 0.7, 
    top_p = 0.9,
)

# Decode Result
print("\n--- Qwen Thinking (Qwen 2.5-3B) ---")
# Strip prompt for cleaness
print(tokenizer.decode(outputs[0]).split("Deconstruction Analysis:")[-1].replace("<|endoftext|>", ""))

Example Output:

--- Qwen Thinking (Qwen 2.5-3B) ---
 <reason>
Diversify across: Equities (Equity ETFs/PE Funds) 20%, Fixed Income (Bonds/Yields) 30%, Cash/Currency 10%, Private Equity/Liquidity Notes 20%, Real Estate 10%. Hedge Funds/Options are 'Unconventional' and risky; keep small.
</reason>
<exception>
The 'Risk-Free Rate' is not zero. Bond prices are falling due to rising interest rates. Cash sits in negative interest. Private equity is expensive and illiquid. The 60-40 portfolio (Stocks vs Bonds) is obsolete in a floating rate world. Unconventional assets like Options/Crypto might offer the needed returns to outperform a boring vanilla portfolio.
</exception>
<tension>
The tension is between *Traditional Diversification* (ETFs/Bonds) and *High-Return Asset Classes* (PE/Private).
</tension>
<categorization>
Keep *Core* in *Assets That Pay* (Bonds/Cash). Invest *Aggressively* in *Liquidity* (PE/Private equity). Don't touch the 'Option Pricing' unless you have the math creds.
</categorization>
<deconstruction>
The goal is 'Net Alpha.' Traditional pools are full. To get alpha, you must either find a hidden risk premium (Value/Equity) or find a productivity advantage (Talent/Process). The complex structures are often just spreadsheets trying to be hedge funds.
</deconstruction>
<conclusion>
A modern hedge fund should prioritize liquid alternative assets like private equity and real estate over traditional bonds/cash, reallocating capital based on the evolving risk-free rate environment.
</conclusion><|im_end|>

Example Training

python

from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import FastLanguageModel, is_bfloat16_supported
from datasets import load_dataset
import torch

# --- 1. CONFIGURATION ---
model_name = "RinKana/Qwen2.5-3B-Deconstruct-V2.4-Merged-v2"


max_seq_length = 4096
dtype = None 
load_in_4bit = True 

# --- 2. LOAD MODEL ---
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = model_name,
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)

# LoRA config
model = FastLanguageModel.get_peft_model(
    model,
    r = 16,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj"],
    lora_alpha = 16,
    lora_dropout = 0,
    bias = "none",
    use_gradient_checkpointing = "unsloth",
    random_state = 3407,
)

# --- 3. FORMATTING FUNCTION (V2 - DECONSTRUCTIONIST) ---
v2_system_prompt = """You are a Deconstructionist Analyst.
User Question: {}
Deconstruction Analysis: {}"""

EOS_TOKEN = tokenizer.eos_token

def formatting_prompts_func(examples):
    questions = examples["Question"]
    reasonings = examples["Reasoning"] 
    texts = []
    for question, reasoning in zip(questions, reasonings):
        text = v2_system_prompt.format(question, reasoning) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts, }

# Load Dataset V3 - 219 dataset
dataset_file = "RinKana/tetralemma-reasoning-dataset-v4" 
dataset = load_dataset(dataset_file, split="train")
dataset = dataset.map(formatting_prompts_func, batched = True)

# --- 4. TRAINING ---
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 2,
    packing = False,
    
    args = TrainingArguments(
        per_device_train_batch_size = 4,
        gradient_accumulation_steps = 2,
        warmup_steps = 5,
        num_train_epochs = 10, 
        learning_rate = 2e-4,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        report_to = "wandb",
        disable_tqdm = False,
    ),
)

# --- 5. START TRAINING ---
print(f"🚀 Starting Eksperimen V2 on {model_name}...")
trainer_stats = trainer.train()

print("✅ Training Done!!!")

Uploaded finetuned model

Developed by: RinKana
License: apache-2.0
Finetuned from model : unsloth/qwen2.5-3b-instruct-bnb-4bit

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.