92 lines
3.3 KiB
Markdown
92 lines
3.3 KiB
Markdown
---
|
|
library_name: transformers
|
|
base_model: meta-llama/Meta-Llama-3-8B-Instruct
|
|
license: llama3
|
|
datasets:
|
|
- MattCoddity/dockerNLcommands
|
|
language:
|
|
- en
|
|
tags:
|
|
- lora
|
|
- docker
|
|
- text-generation
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# llama-3-docker-ft
|
|
|
|
LoRA fine-tune of `meta-llama/Meta-Llama-3-8B-Instruct` that translates natural-language
|
|
requests into Docker CLI commands. Merged adapter weights (base + LoRA), not an
|
|
adapter-only checkpoint.
|
|
|
|
This is a learning-lab artifact ([source notebook and writeup](https://github.com/Fakorede/llm-alignment-lab/tree/main/01_lora_sft)),
|
|
not a production model. Treat it as a first fine-tuning exercise, not a benchmarked release.
|
|
|
|
## Model Details
|
|
|
|
- **Base model:** `meta-llama/Meta-Llama-3-8B-Instruct`, loaded in 8-bit (`BitsAndBytesConfig(load_in_8bit=True)`)
|
|
- **Fine-tuning method:** LoRA (`peft`), `r=16`, `lora_alpha=32`, dropout `0.05`,
|
|
targeting `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
|
|
(41.9M trainable params, 0.52% of 8.07B total)
|
|
- **Merged:** LoRA adapter merged into the base weights before push (`merge_and_unload()`)
|
|
- **License:** inherits the [Llama 3 Community License](https://llama.meta.com/llama3/license/) from the base model
|
|
|
|
## Training Data
|
|
|
|
[`MattCoddity/dockerNLcommands`](https://huggingface.co/datasets/MattCoddity/dockerNLcommands),
|
|
an instruction/input/output dataset pairing natural-language requests with the
|
|
corresponding Docker CLI command. Split 80/20 train/validation (seed 42).
|
|
|
|
## Training Procedure
|
|
|
|
`transformers.Trainer` + `TrainingArguments`: batch size 2, gradient accumulation 4
|
|
(effective batch size 8), `paged_adamw_8bit`, learning rate `2e-4`, 2 epochs (484 steps),
|
|
fp16, warmup steps 5, weight decay 0.01.
|
|
|
|
### Results
|
|
|
|
| Metric | Value |
|
|
|---|---|
|
|
| Train loss (last logged step) | 0.307 |
|
|
| Train loss (run average) | 0.461 |
|
|
| Eval loss | 0.341 |
|
|
| Train runtime | ~2738s (single A100 80GB) |
|
|
|
|
## Uses
|
|
|
|
**Intended use:** translating short, single-turn natural-language instructions about
|
|
containers/images into a Docker CLI command. Example:
|
|
|
|
\`\`\`python
|
|
import transformers
|
|
import torch
|
|
|
|
pipeline = transformers.pipeline(
|
|
"text-generation",
|
|
model="thefabdev/llama-3-docker-ft",
|
|
model_kwargs={"torch_dtype": torch.bfloat16},
|
|
device_map="auto",
|
|
)
|
|
|
|
messages = [
|
|
{"role": "system", "content": "You are a helpful assistant. Translate this sentence in docker command"},
|
|
{"role": "user", "content": "Display the information of the last 4 containers."},
|
|
]
|
|
|
|
result = pipeline(messages, max_new_tokens=256, temperature=0.25, top_p=1, repetition_penalty=1.2)
|
|
print(result[0]["generated_text"][-1]["content"])
|
|
# docker ps --last 4
|
|
\`\`\`
|
|
|
|
**Out of scope:** general-purpose assistant use, multi-turn conversation, any
|
|
command generation where correctness/safety of the resulting shell command isn't
|
|
independently verified before execution. Generated commands are not validated for
|
|
safety and should not be run against production systems without review.
|
|
|
|
## Limitations
|
|
|
|
- Fine-tuned on a single small, narrow dataset (Docker CLI only), will not generalize
|
|
to other CLIs or general instruction-following.
|
|
- Trained for 2 epochs on ~800 examples; not evaluated against a held-out benchmark
|
|
beyond the validation split loss above.
|
|
- No safety/red-teaming evaluation has been performed on this model. |