Pluto-AI-Labs / Atlas-Coder-2-0.5B

---

Model Description

Atlas-Coder-2-0.5B is the flagship model of the Pluto AI research project by Siddharth N.R. — the second generation of the Atlas-Coder series and the most focused coding model released under the Pluto AI brand to date.

Built on top of Qwen2.5-Coder-0.5B-Instruct , Atlas-Coder-2 is trained exclusively on 50K execution-verified OSS-Instruct samples — real open-source Python functions that have been independently verified to execute correctly. This single-source, high-purity data strategy maximizes alignment with HumanEval+ and MBPP+ benchmark formats while keeping the training signal clean and consistent.

Unlike V1 which trained from the base model and used a 4-source mixture, Atlas-Coder-2 starts from an instruction-tuned foundation and specializes it further on execution-verified code. The result is a sharper, more reliable code generator with a lower hallucination rate on self-contained Python tasks.

Research Goal: Demonstrate that a sub-500M parameter model, fine-tuned exclusively on execution-verified code in a single Kaggle session, can match or exceed the coding performance of officially released instruct variants and outperform models up to 3× its parameter count.

---

📊 Benchmarks

Competitor scores from official technical reports.

---

⚡ Edge Performance: Tested locally on an M2 MacBook Air (8GB RAM) using LM Studio with the F16 GGUF . Achieved 75 tokens/second generation speed. Because the model uses native FP16 precision, it bypasses quantization overhead and fully utilizes Apple Silicon's Metal FP16 vector cores. Zero GPU required.

GGUF Quantizations

https://huggingface.co/Siddh07ETH/Atlas-Coder-2-0.5B-GGUF

What Changed from V1

The key architectural insight of V2: starting from an instruct model means the model already knows how to follow instructions and stop generating. V2 doesn't need to re-learn conversation structure — it only needs to deepen its Python code generation capability. This allows a single clean epoch on a smaller, higher-quality dataset to outperform a longer multi-epoch run on a noisier mixture.

---

Training Details

---

Training Data

Why single-source? The V1 multi-dataset mixture introduced noise from TACO (competitive programming verbosity) and CodeFeedback (multi-turn debug style), both of which poorly align with HumanEval+ single-function completion format. V2 eliminates this noise entirely. The OSS-Instruct exec-filtered dataset is already the highest-ROI data source for HumanEval+ performance — using 50K samples of it exclusively produces a cleaner gradient signal than mixing 80K samples of heterogeneous quality.

---

Key Design Decisions

1. Instruct base = faster convergence Starting from Qwen2.5-Coder-0.5B-Instruct means the ChatML format, stop-token behavior, and instruction-following discipline are already in place. The model only needs to deepen code generation quality — not learn conversation structure from scratch. This makes 1 epoch sufficient where V1 needed 3.

2. Execution-verified data only Every training sample in OSS-Instruct exec-filter-50k has been independently run and verified to produce correct output. This eliminates a significant noise source that affects most open-source fine-tuning datasets: plausible-looking but incorrect code completions that silently degrade model performance on pass@1 metrics.

3. r=32 LoRA for speed without sacrificing quality At the 0.5B parameter scale, r=64 provides diminishing returns over r=32 while doubling the LoRA parameter count and training time. The r=32 configuration trains ~40% faster on the T4, allowing full training within a single Kaggle 9-hour session without checkpoint recovery.

4. Single Kaggle session design The entire pipeline — install → load → data → train → merge → upload → GGUF — is designed to complete within a single 9-hour Kaggle session. The 3-layer checkpoint recovery system (local → HuggingFace Hub → fresh start) handles session interruptions automatically when they occur.

---

Usage

Basic Inference

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Siddh07ETH/Atlas-Coder-2-0.5B",
    torch_dtype=torch.float16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Siddh07ETH/Atlas-Coder-2-0.5B")

messages = [
    {
        "role": "system",
        "content": "You are a helpful coding assistant."
    },
    {
        "role": "user",
        "content": "Write a Python function to find all prime numbers up to n using the Sieve of Eratosthenes."
    }
]

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

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.2,
        do_sample=True,
        top_p=0.9,
        repetition_penalty=1.1,
    )

response = tokenizer.decode(
    output[0][inputs.input_ids.shape[1]:],
    skip_special_tokens=True
)
print(response)

Low Memory Inference (4-bit)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
model = AutoModelForCausalLM.from_pretrained(
    "Siddh07ETH/Atlas-Coder-2-0.5B",
    quantization_config=quant_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Siddh07ETH/Atlas-Coder-2-0.5B")

GGUF (Ollama / LM Studio / llama.cpp)

GGUF quantizations for CPU inference are available at:

Siddh07ETH/Atlas-Coder-2-0.5B-GGUF

Runs at 40+ tokens/second on a laptop CPU. No GPU required.

---

Recommended Generation Settings

---

Important: Benchmark Evaluation Setup

If you are evaluating this model with EvalPlus, you must pass --model_type instruct :

python -m evalplus.evaluate   --model Siddh07ETH/Atlas-Coder-2-0.5B   --dataset humaneval   --backend hf   --model_type instruct   --greedy

Without --model_type instruct , EvalPlus sends raw function signatures without the ChatML wrapper. This causes the model to score near 0% — which is an evaluation configuration error, not a reflection of model quality. The model was trained exclusively on ChatML-formatted prompts and will not respond meaningfully to bare code signatures.

---

Model Lineage

Atlas-Coder-2 is the second release in the Atlas-Coder series under Pluto AI. Each version refines the strategy based on lessons from the previous run.

---

Related Models

---

Limitations

• Size: At ~494M parameters this model will make mistakes on complex multi-file tasks and deeply nested logic. Always verify generated code before running it in production.
• Context length: Trained on sequences up to 1024 tokens. Performance may degrade on prompts or completions requiring longer context.
• Language bias: Optimized primarily for Python. Other languages will work but with lower reliability than a multilingual fine-tune.
• Single-domain training: Trained entirely on OSS-Instruct data. May underperform on highly domain-specific code (e.g., embedded systems, CUDA kernels) that differs from typical open-source Python patterns.
• Research only: Not intended for production deployment without further evaluation and safety testing.

---

Author

Siddharth N.R Graduated B.Tech — AI & Data Science Pluto AI Research

---

Citation

@misc{atlascoder2_2026,
  author    = {Siddharth N.R.},
  title     = {Atlas-Coder-2-0.5B: Execution-Verified QLoRA Fine-Tuning from an Instruct Base for Sub-1B Code Generation},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/Siddh07ETH/Atlas-Coder-2-0.5B}
}

---

License

Apache 2.0 — see LICENSE . Base model Qwen2.5-Coder-0.5B-Instruct is also Apache 2.0.