---
base_model: unsloth/Qwen2.5-72B-Instruct-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
- trl
- resume
- career-coach
license: other
language:
- en
---

# 🧠 The Imaginator: Magnum-72B-Career-Strategist

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)

- **Developed by:** jeff-calderon
- **Base Model:** [unsloth/Qwen2.5-72B-Instruct-bnb-4bit](https://huggingface.co/unsloth/Qwen2.5-72B-Instruct-bnb-4bit)
- **Finetuned from model:** unsloth/Qwen2.5-72B-Instruct-bnb-4bit
- **Fine-tuning Framework:** Unsloth / QLoRA

## ⚠️ License & Usage Warning
**License:** **Tongyi Qianwen License Agreement (Research Only / Non-Commercial)**

This model is derived from Qwen-72B. Users must comply with the original Alibaba Cloud Tongyi Qianwen License Agreement. This model is intended for **research and educational purposes only**.

---

## 🎯 The Vision
We set out to build **"The Imaginator"**—not just a generic resume writer, but a high-level **Career Strategist**. 

Most AI resume tools simply fix grammar. The Imaginator is designed to perform cognitive reasoning: it takes "lazy notes" or an "outdated resume," analyzes a specific target job (e.g., *Java Developer* pivoting to *DevOps*), and strategically reframes the candidate's experience to bridge skill gaps without fabricating history.

## 🏗️ The "Trinity" Dataset Strategy
To achieve this level of reasoning, we rejected standard freelance datasets (which often sound like sales pitches). Instead, we engineered a custom "Trinity Dataset" of **7,417 high-quality records** via a local data factory on an RTX 4080:

### 1. Type A: The Stylist (Tone & Impact)
*   **Goal:** Master professional, metric-driven business English.
*   **Input:** Weak, passive bullet points.
*   **Output:** Powerful "STAR" method achievements (Situation, Task, Action, Result).
*   **Source:** Mined 2,000 real resume bullets and utilized Grok to inject industry-standard metrics.

### 2. Type B: The Strategist (Logic & Pivoting)
*   **Goal:** Strategic Reframing.
*   **Input:** A complex JSON payload containing Candidate Context + Target Job + Identified Skill Gaps.
*   **Output:** A rewritten experience section that "bridges the gap" using transferable skills.
*   **Method:** Simulated 1,000 career pivot scenarios (e.g., *Frontend Dev* $\to$ *Full Stack*) using Perplexity/Grok to ensure market accuracy.
*   **Safety:** Rigorously filtered to ensure the model **never** invents fake job titles or promotions.

### 3. Type C: The Creator (Synthesis from Chaos)
*   **Goal:** Structuring unstructured data.
*   **Input:** "Lazy" user brain dumps (lowercase, no formatting, typos).
*   **Output:** Fully formatted, perfectly structured resume sections.
*   **Source:** We used a "Ruiner Script" on 3,000 high-quality resumes to reverse-engineer them into lazy text messages, teaching the model how to reconstruct them.

## 🚀 Capabilities & Performance

This model was fine-tuned on an A100 GPU using Unsloth. It excels at:
*   **Format Synthesis:** Turning raw text into polished documents.
*   **Strategic Pivoting:** Rewriting experience to target specific roles.
*   **Hallucination Control:** Trained specifically *not* to invent fake job titles to fill gaps.

### Inference Example

**Input (Lazy User):**
> "i worked at amazon as a warehouse guy... hit rates... trained new people"

**Imaginator Output:**
> **Logistics Associate** | Amazon
> *Packed products in a timely manner and consistently met or exceeded productivity rates. Trained and mentored new employees on safety protocols and packing procedures, improving team efficiency.*

## 💻 How to Use (Unsloth)

```python
from unsloth import FastLanguageModel
import torch

max_seq_length = 8192
dtype = None
load_in_4bit = True

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "jeff-calderon/Magnum-72B-Imaginator-LoRA", # Your model name here
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are a professional resume writer. Convert the user's raw notes into a polished Experience section."},
    {"role": "user", "content": "managed a team of 5 sales guys. we hit 1m in revenue."}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

outputs = model.generate(input_ids=inputs, max_new_tokens=512, use_cache=True, temperature=0.3)
print(tokenizer.batch_decode(outputs[0], skip_special_tokens=True))