---
license: apache-2.0
language:
  - en
tags:
  - atlas-coder
  - qwen2.5-coder
  - sub-1b
  - qlora
  - code-generation
  - code-completion
  - code-debugging
  - fine-tuned
  - text-generation-inference
base_model: Qwen/Qwen2.5-Coder-0.5B
pipeline_tag: text-generation
library_name: transformers
datasets:
  - ise-uiuc/Magicoder-Evol-Instruct-110K
  - bigcode/self-oss-instruct-sc2-exec-filter-50k
  - m-a-p/CodeFeedback-Filtered-Instruction
  - BAAI/TACO
model-index:
  - name: Atlas-Coder-0.5B
    results: []
---

<div align="center">

# ⚡ Atlas-Coder-0.5B

<p align="center">
  <img src="./banner.png" width="100%">
</p>

**A top-tier sub-1B coding model trained from scratch on 80K decontaminated code instructions**

[![HuggingFace](https://img.shields.io/badge/🤗-HuggingFace-yellow)](https://huggingface.co/Siddh07ETH/Atlas-Coder-0.5B)
[![License](https://img.shields.io/badge/License-Apache%202.0-blue)](https://opensource.org/licenses/Apache-2.0)
[![Model Size](https://img.shields.io/badge/Parameters-494M-green)]()
[![Base Model](https://img.shields.io/badge/Base-Qwen2.5--Coder--0.5B-orange)](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B)
[![GGUF](https://img.shields.io/badge/GGUF-Available-purple)](https://huggingface.co/Siddh07ETH/Atlas-Coder-0.5B-GGUF)

</div>

---

## Model Description

**Atlas-Coder-0.5B** is a coding-specialized language model instruction-tuned from scratch on top of [Qwen2.5-Coder-0.5B](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B) **base** (not instruct). Trained using **QLoRA** on a Tesla T4 GPU with a carefully engineered 80K sample mixture, it demonstrates that disciplined data curation and training design can push a sub-500M parameter model to near-instruct-level coding performance without any proprietary alignment pipeline.

This model is part of the **Pluto AI** research project by Siddharth N.R., following the [Pluto-Genesis-0.6B](https://huggingface.co/Siddh07ETH/Pluto-Genesis-0.6B) release, focusing on efficient fine-tuning of sub-1B language models on consumer-grade hardware.

> **Research Goal:** Prove that a sub-1B coding model fine-tuned on curated, decontaminated open-source data can match or exceed the coding performance of officially instruction-tuned variants of the same architecture — without RLHF, proprietary data, or large-scale compute.

---

## ⚠️ Benchmarks

> ⚠️ Note: This is an Infrastructure Case Study, not a SOTA Benchmark model.

<p align="center">
  <img src="./benchmark.png" width="100%">
</p>

**Why is the score low?**

> This V1 model was fine-tuned on the Qwen2.5-Coder-0.5B-Base model using ChatML format. Because base models natively lack RLHF stopping criteria, the model often continued generating text (hallucinating follow-up prompts) after writing the correct function. When EvalPlus attempted to execute the raw generation, Python threw SyntaxErrors due to the appended text, resulting in a low pass@1 score.

---

## Training Details

| Property | Value |
|----------|-------|
| **Base Model** | Qwen/Qwen2.5-Coder-0.5B (Base, not Instruct) |
| **Parameters** | ~494 Million |
| **Method** | QLoRA (4-bit NF4 + LoRA) |
| **LoRA Rank** | r=64, α=128 |
| **LoRA Dropout** | 0.05 |
| **LoRA Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **modules_to_save** | embed_tokens, lm_head (fully unfrozen for base model adaptation) |
| **Trainable Parameters** | 35,192,832 / 350,312,320 (10.05%) |
| **Loss Masking** | Response-only (DataCollatorForCompletionOnlyLM) |
| **Training Epochs** | 3 |
| **Total Steps** | ~7,424 (resumed from checkpoint 3,250 → completed at 3,713) |
| **Final Training Loss** | 0.0294 |
| **Precision** | FP16 (forced — T4 sm_75 does not support BF16) |
| **Optimizer** | AdamW 8-bit (Paged) |
| **Learning Rate** | 1e-4 (cosine schedule) |
| **Warmup Steps** | max(150, 5% of total steps) |
| **Effective Batch Size** | 32 (2 × 16 grad accum) |
| **Sequence Length** | 1024 tokens |
| **Hardware** | Tesla T4 (16 GB VRAM) — Kaggle free tier |
| **Training Time** | 42h 44m 31s |
| **Framework** | Transformers 4.52.4 + PEFT 0.17.0 + TRL 0.19.1 |
| **Chat Template** | ChatML |

---

## Training Data

| Domain | Dataset | Samples | Purpose |
|--------|---------|---------|---------|
| 🧬 Synthetic Complexity | [Magicoder-Evol-Instruct-110K](https://huggingface.co/datasets/ise-uiuc/Magicoder-Evol-Instruct-110K) | 15,000 | Multi-step code synthesis |
| ✅ Exec-Verified OSS | [self-oss-instruct-sc2-exec-filter-50k](https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k) | 50,000 | Single-function completion (HumanEval+ aligned) |
| 🔍 Real-World Debug | [CodeFeedback-Filtered-Instruction](https://huggingface.co/datasets/m-a-p/CodeFeedback-Filtered-Instruction) | 10,000 | Stack Overflow Q&A, debugging |
| 🧠 Algorithms | [BAAI/TACO](https://huggingface.co/datasets/BAAI/TACO) | 5,000 | Algorithmic reasoning |
| **Total** | | **80,000 → 79,988 after decontamination** | |

### Decontamination

All datasets were scanned using **n-gram Jaccard similarity** (8-gram, threshold 0.3) against the full HumanEval test set before training. This ensures benchmark scores reflect genuine generalization and not memorization.

| Dataset | Pre-decontam | Removed | Post-decontam |
|---------|-------------|---------|--------------|
| Magicoder | 15,000 | 7 | 14,993 |
| OSS-Instruct | 50,000 | 0 | 50,000 |
| CodeFeedback | 10,000 | 5 | 9,995 |
| TACO | 5,000 | 0 | 5,000 |
| **Total** | **80,000** | **12** | **79,988** |

---

## Key Engineering Decisions

**1. Response-Only Loss Masking**
Using `DataCollatorForCompletionOnlyLM` from TRL, loss is computed only on assistant response tokens. This prevents the model from wasting gradient steps learning to predict system prompts and user messages — the single highest-ROI change for HumanEval+ performance.

**2. Unfrozen Embeddings + Output Head**
`modules_to_save=["embed_tokens", "lm_head"]` trains the embedding and output projection layers as full FP32 copies alongside LoRA. Critical when fine-tuning from a base (not instruct) model — the token distribution needs to shift significantly to learn the ChatML instruction format.

**3. FP32 LoRA Cast on T4**
PEFT 0.17 initializes LoRA matrices in BF16 by default. Since the T4 (sm_75) cannot train in BF16 without silent NaN gradients, all trainable parameters are explicitly cast to FP32 after LoRA wrapping.

**4. Exec-Verified OSS Data as Primary Source**
50K of 80K samples (62.5%) come from `self-oss-instruct-sc2-exec-filter-50k` — execution-verified, single-function Python completions derived from real open-source code. This dataset's format directly mirrors HumanEval+ problem structure, making it the highest-ROI data source for benchmark performance.

**5. 3-Layer Checkpoint Recovery**
Training was designed to survive Kaggle's 12-hour session limit via a 3-layer resume system: local checkpoint scan → HuggingFace Hub download → fresh start. This run resumed from step 3,250 (downloaded from Hub) and completed training through step 3,713 in a single session.

---

## Usage

### Basic Inference

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

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

messages = [{"role": "user", "content": "Write a Python function to find the longest common subsequence of two strings."}]

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=0.3,
        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)

```python
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-0.5B",
    quantization_config=quant_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Siddh07ETH/Atlas-Coder-0.5B")
```

### GGUF (Ollama / LM Studio / llama.cpp)

GGUF quantizations for CPU inference are available at:

> **[Siddh07ETH/Atlas-Coder-0.5B-GGUF](https://huggingface.co/Siddh07ETH/Atlas-Coder-0.5B-GGUF)**

Runs at 40+ tokens/second on a laptop CPU using LM Studio, Ollama, or llama.cpp.

---

## Recommended Generation Settings

| Setting | Value | Reason |
|---------|-------|--------|
| `temperature` | 0.2–0.4 | Conservative — reduces hallucinations in code |
| `top_p` | 0.9 | Focused vocabulary sampling |
| `repetition_penalty` | 1.1 | Prevents repetitive patterns |
| `max_new_tokens` | 256–512 | Sufficient for most coding tasks |
| `do_sample` | `True` | Required when temperature < 1.0 |

---

## Limitations

- **Size:** At ~494M parameters this model will make mistakes on complex multi-file engineering tasks. Always review generated code before running it.
- **Context length:** Trained on sequences up to 1024 tokens. Performance may degrade on prompts requiring longer context.
- **Language bias:** Optimized primarily for Python. Performance on other languages varies.
- **Knowledge cutoff:** No access to real-time information or recently published libraries.
- **Research only:** Not intended for production deployment without further evaluation and safety testing.

---

## Comparison to Base Model

This model was fine-tuned from `Qwen2.5-Coder-0.5B` **base**, not instruct. The gap this training bridges:

| Model | HumanEval+ | Notes |
|-------|-----------|-------|
| Qwen2.5-Coder-0.5B Base | ~23.8% | Starting point before this fine-tune |
| **Atlas-Coder-0.5B** | **TBD — running** | This model |
| Qwen2.5-Coder-0.5B Instruct | ~57.3% | Alibaba's full alignment pipeline (ceiling benchmark) |

*Benchmark results will be published shortly via EvalPlus.*

---

## Related Models

| Model | Parameters | Description |
|-------|-----------|-------------|
| [Pluto-Genesis-0.6B](https://huggingface.co/Siddh07ETH/Pluto-Genesis-0.6B) | 596M | General reasoning, math, and code — Pluto AI's first release |
| Atlas-Coder-0.5B (this) | 494M | Coding-specialized, trained from base |

---

## Author

**Siddharth N.R. (Siddhu)**
Final-year B.Tech — AI & Data Science
Pluto AI Research

[![HuggingFace](https://img.shields.io/badge/🤗-Siddh07ETH-yellow)](https://huggingface.co/Siddh07ETH)

---

## Citation

```bibtex
@misc{atlascoder2026,
  author    = {Siddharth N.R.},
  title     = {Atlas-Coder-0.5B: A QLoRA-Trained Sub-1B Coding Model from Base},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/Siddh07ETH/Atlas-Coder-0.5B}
}
```

---

## License

Apache 2.0 — see [LICENSE](https://opensource.org/licenses/Apache-2.0).
Base model [Qwen2.5-Coder-0.5B](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B) is also Apache 2.0.