🌸 PyBlissa-Coder-40M
!! 14.6% SCORE ON HumanEval PASS@1 !!
PyBlissa-Coder-40M is the second model from the PyBlissa-Coder family that mainly supports python coding. Despite its small footprint, 40M parameters, trained on 272M tokens, PyBlissa has achieved an amazing score of 14.6% on the HumanEval dataset; and 4.4% on MBPP dataset (both being benchmark datasets). While scoring such a good number, it's imperfections are also something to be aware of. This model can sometimes generate wrong, inefficient, or broken codes. Though its mostly temperature-dependent.
Benchmarks
| Benchmark | Score | Protocol | Temp |
|---|---|---|---|
| HumanEval pass@1 | 14.6% (24/164) | zero-shot, fenced-code extraction | 0.25 |
| MBPP pass@1 | 4.4% (22/500) | official tests-in-prompt (Austin et al. 2021) | 0.05 |
How PyBlissa compares on HumanEval
| Model | Params | HumanEval pass@1 |
|---|---|---|
| GPT-Neo | 125M | 0.75% |
| CodeParrot-small | 110M | 3.80% |
| PyCodeGPT | 110M | 8.33% |
| PyBlissa-Coder | 40M | 14.6% |
PyBlissa is ~2.75× smaller than CodeParrot-small yet scores roughly 4× higher on HumanEval pass@1, trained on a single consumer GPU.
Model details
| Architecture | Decoder-only transformer (GPT-2 style, nanoGPT lineage) |
| Parameters | 39.9M |
| Layers | 10 |
| Model dim (d_model) | 512 |
| Heads | 8 (head_dim 64) |
| FFN dim (d_ff) | 2048 |
| Context length | 512 tokens |
| Vocab size | 16,000 (custom ByteLevel BPE) |
| Tied embeddings | Yes |
| Precision | trained in bf16, released as F32 GGUF |
| Best val loss | 0.3615 |
Training
| Hardware | 1 × NVIDIA RTX 5080 (16 GB) |
| Training tokens | 272M (train split) |
| Epochs | 5 |
| Optimizer | AdamW (β 0.9/0.95, wd 0.1) |
| LR schedule | cosine, 4e-4 → 4e-5, ~2% warmup |
| Batch size | 48 |
| Total steps | 55,405 |
| Wall-clock time | ~116 min |
Usage
Ollama
ollama run hf.co/Rohanify/PyBlissa-Coder-40M:F32
The repo ships template and params files, so Ollama applies the correct
PROMPT:/CODE: format and sampling defaults automatically — no Modelfile
needed for remote runs.
To run a local GGUF instead:
ollama create pyblissa-40m -f Modelfile
ollama run pyblissa-40m "write a function that checks if a number is prime"
Prompt format
The model was trained on a plain-text wrapper. At inference, the prompt is wrapped as:
PROMPT: {your instruction}
CODE:
The model then emits a fenced ```python code block. (When using Ollama, the
template file does this wrapping for you — just type a plain instruction.)
Recommended sampling
| Parameter | Value |
|---|---|
| temperature | 0.25 – 0.3 |
| top_k | 10 |
| repeat_penalty | 1.25 |
| num_ctx | 512 |
Limitations
PyBlissa is a 40M-parameter model trained primarily for prompt → Python generation. Known limitations:
- It is a small model: it solves short, self-contained functions well but struggles with multi-step or library-heavy tasks.
- It sometimes omits
importstatements for stdlib modules it uses (math,re,hashlib, etc.). - It can occasionally emit a short natural-language preamble before the code block on harder prompts.
- Code explanation and non-Python tasks are out of distribution — it may attempt them, but that is not what it was trained for.
- As with any code model, review and test generated code before running it.
Training data & attribution
This model was trained on the following datasets. Per their licenses, attribution is provided here:
- nvidia/OpenCodeInstruct — CC-BY-4.0 https://huggingface.co/datasets/nvidia/OpenCodeInstruct
- flytech/python-codes-25k — MIT https://huggingface.co/datasets/flytech/python-codes-25k
No OpenAI-derived data was used in training.
License
The model weights are released under Apache-2.0. Note that the training data carries its own licenses (CC-BY-4.0 and MIT, see above), which require attribution as provided.
@misc{pyblissa2026,
title = {PyBlissa-Coder-40M: A from-scratch Python code model},
author = {Rohan},
year = {2026},
howpublished = {\url{https://huggingface.co/Rohanify/PyBlissa-Coder-40M}}
}