🧠 Qwen3.5-9B-Trace-Inversion

This repository contains a fully merged, standalone model based on Qwen/Qwen3.5-9B.

It was fine-tuned on the Trace-Inversion-1K-Preview dataset to master a specific technique called Trace Inversion.

🎯 What does this model do?

Instead of solving a problem from scratch, this model acts as a Reasoning Generator. You provide it with a problem and its known correct final answer, and it reconstructs the step-by-step logical reasoning (<think>...</think>) that connects the two.

This 9B version is significantly more capable than its 4B predecessor and is highly useful for:

Distilling reasoning capabilities from massive models (like DeepSeek V4 or Claude Opus) into smaller 9B architectures.
Data Augmentation: Generating synthetic Chain-of-Thought traces for existing datasets.
Behavioral Cloning: Teaching small models how to structure their thoughts before answering, with better logic retention than smaller models.

🚀 How to use

Since this is a full model, you can load and use it directly with the standard transformers library.

(Note: Qwen3.5 models require explicit eos_token_id settings to prevent infinite <|im_end|> looping. The code below handles this correctly).

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "open-zzrl/Qwen3.5-9B-Trace-Inversion"

# 1. Load Model and Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)
model.eval()

# 2. Prepare the prompt (Question + Known Answer)
problem = "How many trailing zeros are there in 100! (100 factorial)?"
known_answer = "24"

messages = [
    {"role": "user", "content": f"Question: {problem}\n\nAnswer: {known_answer}"}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True, 
    return_tensors="pt"
).to(model.device)

# 3. Generate the reasoning
with torch.no_grad():
    outputs = model.generate(
        inputs, 
        max_new_tokens=2048, 
        temperature=0.6,
        do_sample=True,
        # CRITICAL: Prevents infinite <|im_end|> spam
        eos_token_id=tokenizer.convert_tokens_to_ids("<|im_end|>"),
        pad_token_id=tokenizer.pad_token_id
    )

# 4. Decode and print
generated_text = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(generated_text)

⚠️ Limitations & Caveats

Do not evaluate on standard benchmarks (AIME, MATH-500, HumanEval). These benchmarks were part of the training data. The model will likely score artificially high, not because it learned to generalize, but because it memorized the traces.
It is not a general-purpose solver. This model expects the answer to be provided in the prompt. If you ask it to solve a problem without giving it the answer, it will likely fail.
9B Parameter Behavior. The 9B model shows excellent reasoning capabilities, successfully handling advanced university-level math (e.g., Legendre's formula) and complex algorithmic logic (e.g., Two Pointers). However, it may still occasionally double-count or make logical leaps on extremely tricky combinatorial problems.
Qwen Native Tags. The dataset and this model use Qwen's native <think> and </think> tags.