Skip to content
View Sergasgr's full-sized avatar

Block or report Sergasgr

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Sergasgr/README.md

Sergio Graciá del Cisne

Final-year Computer Science student interested in Machine Learning and AI, focused on fine-tuning and alignment of LLMs. Alongside this, I work as an AI Engineer at Dialapplet, working hands-on with Python/FastAPI services, MLOps infrastructure, Agentic AI, and NLP model training pipelines.

Projects

Three-part exploration of a specialized coding LLM's full lifecycle: align it, compress it, serve it fast.

  • CodeAlign — Post-training Qwen2.5-Coder-7B-Instruct with SFT + DPO, using a composite reward (execution success + code complexity + lint score) instead of binary pass/fail, to avoid reward-hacking toward code that passes tests but is needlessly complex. Covers 8 languages spanning JetBrains' IDE lineup.
  • CodeAlign-Runtime (in development) — Minimal C++/CUDA inference engine to serve the model with low latency: naive CUDA kernels → optimized kernels → INT4 quantization, with a final comparison against llama.cpp planned.
  • CodeAlign-Distillation (future) — Knowledge distillation + QAT to compress CodeAlign's DPO checkpoint into a 5-14x smaller model, closing the loop for CodeAlign-Runtime.

Open Source Contributions

peft — Benchmarked 3 new PEFT methods (BEFT, HiRA, AdaMSS) against LoRA/OFT baselines on the image-gen benchmark, wired up ASA training-callback support in the image-gen benchmark's script run.py (previously unimplemented), and fixed a target-modules matching bug along the way.

Discussion #3522 · PRs: #3607 #3640 #3641 #3661 #3663

Pinned Loading

  1. huggingface/peft huggingface/peft Public

    🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.

    Python 21.6k 2.5k

  2. CodeAlign CodeAlign Public

    Aligning code LLMs via DPO with a composite reward (execution + static analysis) to produce clean, maintainable code — not just code that passes tests.

    Jupyter Notebook

  3. CodeAlign-Runtime CodeAlign-Runtime Public

    Low-latency C++/CUDA inference engine for small code LLMs. Features custom memory-bound optimized GEMV kernels and INT4 quantization for fast autoregressive decoding.

    Python