README.md

---
language: en
license: apache-2.0
library_name: transformers
tags:
- text-generation
- competitive-programming
- code-generation
- cpp
- unsloth
- llama
- trl
base_model: unsloth/llama-3.1-8b-instruct-bnb-4bit
datasets:
- Redhanuman/soltra-codeforces-cpp-elite-10k
---

## Soltra: A Llama-3.1 8B Fine-tune for Elite Competitive Programming
_This isn't your average "hello world" code-gen model._

Ever been stuck on a Codeforces problem and wished you had a buddy who's already grinded thousands of them? That's Soltra. This model was fine-tuned with a very specific mission: to be a high-level thought partner for **elite-level competitive programming (1800-2600 rating)**, focusing exclusively on **C++**.

It's designed to help you break down complex problems, generate solid C++ implementations, and understand the underlying algorithmic patterns.

---

### The Secret Sauce: Curated, No-Fluff Data 🧠

We all know the rule: garbage in, garbage out. The reason Soltra performs well is the data it was trained on. This wasn't a random scrape. I built a custom pipeline to create a dataset of gold-standard solutions.

The model was trained on my **[soltra-codeforces-cpp-elite-10k](https://huggingface.co/datasets/Redhanuman/soltra-codeforces-cpp-elite-10k)** dataset, which was filtered with these strict rules:
* **Problem Rating:** Only problems between **1800 and 2600**. No beginner stuff.
* **Verdict:** Only submissions with a **`verdict: 'OK'`**. This model only learns from code that works.
* **Language:** C++ only, sourced from a massive pool of 3 million submissions.
* **Rich Context:** Each entry includes the problem statement, rating, and tags, teaching the model to connect the problem description to the solution structure.

---

### Under the Hood: The Tech Stack 🛠️

This was a classic solo-dev project running on a tight budget. Here’s what’s powering Soltra:

* **Base Model:** `unsloth/llama-3.1-8b-instruct-bnb-4bit` - A powerful and modern foundation.
* **Fine-tuning:** [Unsloth](https://github.com/unslothai/unsloth) + Hugging Face's TRL for a 2x speed boost and 50%+ less memory usage.
* **Quantization:** 4-bit quantization, making it possible to train and run this on a single free Colab T4 GPU.
* **Hardware:** One motivated developer and one T4 GPU. That's it.

---

### How to Use Soltra 🚀

Alright, enough talk. Here's the boilerplate to get this running. Since these are LoRA adapters, you first load the base model and then apply the fine-tuned weights on top.

```python
from unsloth import FastLanguageModel
import torch

# The original base model
base_model_name = "unsloth/llama-3.1-8b-Instruct-bnb-4bit"

# STEP 1: Load the base model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = base_model_name,
    max_seq_length = 4096,
    dtype = None,
    load_in_4bit = True,
)

# STEP 2: Apply your fine-tuned adapters from the Hub
# This is where you load Soltra's brain
model = FastLanguageModel.from_pretrained(
    model = model,
    model_name = "Redhanuman/soltra-llama-3.1-8b-cpp-adapters", # Your repo on the Hub
)

# --- Now, run inference ---
# The prompt must be in the same format the model was trained on.
prompt = """<|begin_of_text|><|start_header_id|>user<|end_header_id|>

Solve this competitive programming problem by providing a step-by-step thought process and then the final code.

**Problem:** C. Registration System
**Rating:** 1500
**Tags:** data structures, strings, maps

**Problem Statement:**
A new user registration system is being developed. When a new user wants to register, they enter a desired username. If this name is not already in the database, it's added, and the user receives an "OK" message. If the name is already taken, the system appends a number to the name to make it unique. The first time a name is duplicated, it appends '1', the second time '2', and so on. Given a sequence of username registration attempts, output the system's response for each.

**Provide:**
1.  **Thought Process:** A brief explanation of the logic, data structures, and algorithm used.
2.  **C++ Solution:** An efficient and correct solution in C++.<|eot_id|><|start_header_id|>assistant<|end_header_id|>"""

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

# Generate the response
with torch.no_grad():
  outputs = model.generate(**inputs, max_new_tokens=512, use_cache=True)
response = tokenizer.batch_decode(outputs)

# Print the generated part of the response
print(response[0].split("<|start_header_id|>assistant<|end_header_id|>")[1].replace("<|eot_id|>", "").strip())
```
### Intended Use & Limitations ⚠️


* **Use it for:** Getting ideas on difficult problems, understanding common C++ patterns for certain tags (like DP or graphs), and generating boilerplate code for standard algorithms.
* **Don't use it for:** Blindly copy-pasting into a contest. The model might not always produce the most optimal solution, and it might hallucinate on edge cases it hasn't seen.

_Think of Soltra as a highly-skilled coding buddy, not a replacement for your own brain._
---
Built by Redhanuman.