llama31-8b-aurora-chat-v3-gguf/README.md

---
library_name: gguf
license: apache-2.0
language: [en]
base_model: meta-llama/Llama-3.1-8B-Instruct
pipeline_tag: text-generation
tags:
  - aurora
  - alcf
  - hpc
  - intel-gpu
  - oneapi
  - sycl
---

# Llama-3.1-8B-Aurora-Chat v3

🏆 **Best Aurora chat model in our zoo (eval 2.80/5, +59% over base).**

LoRA fine-tune of [`meta-llama/Llama-3.1-8B-Instruct`](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) specialized for the
[**ALCF Aurora supercomputer**](https://docs.alcf.anl.gov/aurora/) (Intel Xeon Sapphire
Rapids + Intel GPU Max 1550 / Ponte Vecchio, oneAPI / SYCL, PBS Pro).

Off-the-shelf code-LLMs hallucinate Aurora specifics — they suggest `nvcc` instead of
`icpx -fsycl`, `srun` / `aprun` instead of `mpiexec`, NERSC's `/global/cfs` instead of
`/lus/flare`, and CUDA device strings instead of `xpu`. This adapter teaches the base
model the actual Aurora toolchain, file system layout, scheduler conventions, and
recommended PyTorch/TensorFlow/SYCL idioms.

## Model summary

| | |
|---|---|
| **Base model** | [`meta-llama/Llama-3.1-8B-Instruct`](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
| **Format** | GGUF, f16 — single file, llama.cpp / Ollama / LM Studio compatible |
| **Fine-tuning** | LoRA (PEFT) — r=32, α=64, dropout 0.0, 2 epochs |
| **Optimizer** | AdamW fused, lr 2e-4 cosine, warmup 3%, batch 1 × grad-accum 8 |
| **Precision / seq-len** | bf16, 1,536 tokens |
| **Training data** | [`aurora-docs-distill-multirank`](https://github.com/SIslamMun/Generator/tree/aurora-datasets-2026-04-30/datasets/aurora/iter2/data/training/A) — 4,495 ChatML rows |
| **Train loss (final)** | 0.6224 |
| **Hardware** | 1 Aurora PVC tile (1/12 of a node, 64 GB HBM), IPEX + PyTorch 2.10 XPU backend |
| **Eval (53-Q Aurora, 0–5)** | **2.80 / 5**   *(base 1.76, +59.1%)* |

## Quick start

**On Aurora** (PVC GPU, SYCL llama.cpp build) — interactive PBS session:
```bash
# 1. Grab a debug node
qsub -I -A <project> -q debug -l select=1,walltime=01:00:00,filesystems=home:flare

# 2. Load the toolchain
module load frameworks
source /lus/flare/projects/<project>/scripts/env.sh   # or your own oneAPI setup
export ONEAPI_DEVICE_SELECTOR=level_zero:gpu

# 3. Download to flare (NOT $HOME — quota is small)
hf download shazzadulimun/llama31-8b-aurora-chat-v3-gguf --local-dir /lus/flare/projects/<project>/models/aurora-chat-v3

# 4. Run on a single PVC tile
/path/to/llama.cpp/build_sycl/bin/llama-cli \
    -m /lus/flare/projects/<project>/models/aurora-chat-v3/*.gguf \
    -ngl 999 -sm none --temp 0.0 -cnv \
    -p "How do I launch one MPI rank per GPU tile on Aurora?"
```

**Anywhere else** (laptop, workstation, any GPU):
```bash
hf download shazzadulimun/llama31-8b-aurora-chat-v3-gguf --local-dir ./model
./llama-cli -m ./model/*.gguf -ngl 999 --temp 0.0 -cnv
```

Or **Ollama / LM Studio**: `ollama run hf.co/shazzadulimun/llama31-8b-aurora-chat-v3-gguf`


## Training data

Distilled from `openai/gpt-oss-120b on ALCF Sophia (vLLM)` over 416 cleaned chunks of
[`docs.alcf.anl.gov/aurora`](https://docs.alcf.anl.gov/aurora/). 4,495
training rows + 562 validation rows in ChatML format with embedded
chain-of-thought (`**Reasoning:**` / `**Answer:**`).

**Broad coverage, parallel-rank distillation.** 20 worker ranks each took a *disjoint* slice (~21 chunks) of the cleaned `docs.alcf.anl.gov/aurora` corpus and asked the teacher for chain-of-thought QA pairs. Disjoint slicing maximizes phrasing diversity (each rank sees fresh context) while still covering every chunk exactly once.

Full corpus + reproduction scripts:
[**SIslamMun/Generator @ aurora-datasets-2026-04-30**](https://github.com/SIslamMun/Generator/tree/aurora-datasets-2026-04-30/datasets/aurora/iter2/data/training/A).

## Evaluation

53-question Aurora-domain holdout (programming models, ML/AI, systems/ops, debugging).
Judged by `gpt-oss-120b` on a 0–5 scale.

| Model | Avg | Δ vs. base |
|---|---|---|
| Llama-3.1-8B-Aurora-Chat v3 (`-A` data) — best | **2.80** | +59% |
| Llama-3.1-8B-Aurora-Ops v3 | 2.31 | +31% |
| Llama-3.1-8B-Aurora-Chat v1 (`-B` data, single-rank ablation) | 2.45 | +39% |
| Llama-3.1-8B-Aurora-ML v3 | 2.13 | +21% |
| Llama-3.1-8B-Aurora-Coder v3 | 1.97 | +12% |
| `meta-llama/Llama-3.1-8B-Instruct` (base) | 1.76 | — |

Closed frontier models (gpt-4o, claude-sonnet-4-5, the gpt-oss-120b teacher) score
3.6–4.1 on the same holdout — the goal here isn't to beat them, it's to distill enough
Aurora knowledge into a small open model that runs on a single PVC tile.

## Limitations

- **Synthetic-data biases.** Teacher (`gpt-oss-120b`) can confabulate plausible-looking
  but incorrect commands. Treat outputs as a verifiable first draft, not authoritative.
- **Doc snapshot is fixed at 2026-04-29.** Module versions, queue names, and APIs change
  — anything published after that date isn't reflected here.
- **Aurora-only.** Specifics (`/lus/flare`, `xpu`, PBS queues) won't transfer to Frontier,
  Polaris, or other systems.
- **Use temperature ≤ 0.1** for technical answers; higher temps invite invented flag names
  and paths.

## Citation

```bibtex
@misc{aurora-llms-2026,
  title  = { Llama-3.1-8B-Aurora-Chat v3 },
  author = { Islam Mun, Shazzadul },
  year   = { 2026 },
  url    = { https://huggingface.co/shazzadulimun/llama31-8b-aurora-chat-v3-gguf },
  note   = { LoRA fine-tune of Llama-3.1-8B-Instruct; data distilled from gpt-oss-120b on docs.alcf.anl.gov/aurora }
}
```

## License

Apache-2.0 for the adapter weights and synthetic training data. Source corpus is public
ALCF user documentation. Base model retains its own license — see
[`meta-llama/Llama-3.1-8B-Instruct`](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct).