Instructions to use antoninbrthn/skill-neologisms-qwen2.5-digitseq-base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use antoninbrthn/skill-neologisms-qwen2.5-digitseq-base with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B")
model = PeftModel.from_pretrained(base_model, "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base")

Transformers

How to use antoninbrthn/skill-neologisms-qwen2.5-digitseq-base with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="antoninbrthn/skill-neologisms-qwen2.5-digitseq-base")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("antoninbrthn/skill-neologisms-qwen2.5-digitseq-base", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use antoninbrthn/skill-neologisms-qwen2.5-digitseq-base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/antoninbrthn/skill-neologisms-qwen2.5-digitseq-base

SGLang

How to use antoninbrthn/skill-neologisms-qwen2.5-digitseq-base with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "antoninbrthn/skill-neologisms-qwen2.5-digitseq-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use antoninbrthn/skill-neologisms-qwen2.5-digitseq-base with Docker Model Runner:
```
docker model run hf.co/antoninbrthn/skill-neologisms-qwen2.5-digitseq-base
```

Model Card for Model ID

This is a saved checkpoint from fine-tuning a Qwen/Qwen2.5-0.5B model via LoRA for the digit-sequence transformation task the paper "Skill Neologisms: Towards Skill-based Continual Learning" (ICML 2026).

Model Details

Model Description

This checkpoint was fine-tuned on a synthetic digit-sequence transformation task (see paper for full details).

Finetuned from model [optional]: Qwen/Qwen2.5-0.5B

Model Sources

Repository: Official code for the paper "Skill Neologisms: Towards Skill-based Continual Learning"
Paper: Skill Neologisms: Towards Skill-based Continual Learning

Uses

Please refer to the official repo for example use.

Training Details

The LoRA adapter is trained over two phases with the configuration below (please refer to the paper for more details).

Training configuration.
Parameter	Phase 1	Phase 2
Base Model	Qwen/Qwen2.5-0.5B
PEFT Method	LoRA (r=32, α=32)
Target Modules	q, k, v, o, gate, up, down
Training Samples	100,000	500,000
Test Samples	500	500
Operations per Sample	1	1–3
Epochs	3	3
Batch Size	64	64
Learning Rate	2e-4	2e-4
Warmup Steps	500	500
Operations	`[ASC]`, `[DESC]`, `[ADD]`, `[SUB]`, `[POL]`, `[REV]`, `[ID]`
Sequence Lengths	2, 3, 4, 6, 8 (held-out: 5, 7, 9)
Held-out 3-op combinations	--	25%

Technical Specifications

Hardware

This model was trained on a NVIDIA RTX 6000 GPU (48GB VRAM).

Citation

BibTeX:

@article{berthon2026skill,
  title={Skill Neologisms: Towards Skill-based Continual Learning},
  author={Berthon, Antonin and Astorga, Nicolas and van der Schaar, Mihaela},
  journal={arXiv preprint arXiv:2605.04970},
  year={2026}
}

Model Card Contact

Antonin Berthon (berthon/dot\antonin/at\gmail/dot\com)

Framework versions

transformers 5.0.0rc1
peft 0.17.1
trl 0.26.1

Downloads last month: 212

Model tree for antoninbrthn/skill-neologisms-qwen2.5-digitseq-base

Base model

Qwen/Qwen2.5-0.5B

Adapter

(409)

this model

Paper for antoninbrthn/skill-neologisms-qwen2.5-digitseq-base

Skill Neologisms: Towards Skill-based Continual Learning

Paper • 2605.04970 • Published 24 days ago