Instructions to use Redhanuman/soltra-llama-3.1-8b-cpp-adapters with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Redhanuman/soltra-llama-3.1-8b-cpp-adapters with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Redhanuman/soltra-llama-3.1-8b-cpp-adapters")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Redhanuman/soltra-llama-3.1-8b-cpp-adapters", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Redhanuman/soltra-llama-3.1-8b-cpp-adapters with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Redhanuman/soltra-llama-3.1-8b-cpp-adapters"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Redhanuman/soltra-llama-3.1-8b-cpp-adapters",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Redhanuman/soltra-llama-3.1-8b-cpp-adapters

SGLang

How to use Redhanuman/soltra-llama-3.1-8b-cpp-adapters with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Redhanuman/soltra-llama-3.1-8b-cpp-adapters" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Redhanuman/soltra-llama-3.1-8b-cpp-adapters",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Redhanuman/soltra-llama-3.1-8b-cpp-adapters" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Redhanuman/soltra-llama-3.1-8b-cpp-adapters",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Studio new

How to use Redhanuman/soltra-llama-3.1-8b-cpp-adapters with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Redhanuman/soltra-llama-3.1-8b-cpp-adapters to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Redhanuman/soltra-llama-3.1-8b-cpp-adapters to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Redhanuman/soltra-llama-3.1-8b-cpp-adapters to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="Redhanuman/soltra-llama-3.1-8b-cpp-adapters",
    max_seq_length=2048,
)

Docker Model Runner
How to use Redhanuman/soltra-llama-3.1-8b-cpp-adapters with Docker Model Runner:
```
docker model run hf.co/Redhanuman/soltra-llama-3.1-8b-cpp-adapters
```

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 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 + 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.

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.

Downloads last month: -; Downloads are not tracked for this model. How to track

Redhanuman
/

soltra-llama-3.1-8b-cpp-adapters