Instructions to use anicka/nla-qwen3-4b-universal-ar with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use anicka/nla-qwen3-4b-universal-ar with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B")
model = PeftModel.from_pretrained(base_model, "anicka/nla-qwen3-4b-universal-ar")

Transformers

How to use anicka/nla-qwen3-4b-universal-ar with Transformers:

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

pipe = pipeline("text-generation", model="anicka/nla-qwen3-4b-universal-ar")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("anicka/nla-qwen3-4b-universal-ar", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use anicka/nla-qwen3-4b-universal-ar with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "anicka/nla-qwen3-4b-universal-ar"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "anicka/nla-qwen3-4b-universal-ar",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/anicka/nla-qwen3-4b-universal-ar

SGLang

How to use anicka/nla-qwen3-4b-universal-ar 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 "anicka/nla-qwen3-4b-universal-ar" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "anicka/nla-qwen3-4b-universal-ar",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "anicka/nla-qwen3-4b-universal-ar" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "anicka/nla-qwen3-4b-universal-ar",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use anicka/nla-qwen3-4b-universal-ar with Docker Model Runner:
```
docker model run hf.co/anicka/nla-qwen3-4b-universal-ar
```

NLA Activation Reconstructor — Qwen3 4B, universal multi-layer

LoRA adapter that reads a natural-language description of an activation and reconstructs the activation vector — the audit half of the NLA pair. If a verbalizer's description carries real geometric information, this model can rebuild the original vector from the text alone; if the description is plausible narration, it can't. It is also the reward model for the GRPO-refined verbalizer.

Part of the nla-at-home project. Companions: AV (SFT), AV (GRPO).

Architecture: self-layer readout (`lora_sl`)

No value heads. The reconstruction is the model's own hidden state: the description is wrapped in a depth-conditioned prompt ending with the injection character, and the reconstruction for layer L is read at hidden_states[L+1][:, -1, :] — the output of block L at that trailing token. One adapter serves all 36 layers.

Prompt (frozen interface, nla_lib.AR_TEMPLATE_DEPTH_SL):

Summary of the following text from depth {depth}%: <text>{explanation}</text> <summary>㈎

Left padding, add_special_tokens=False, max length 512. Load with nla_lib.load_ar_lora_sl(...) / score with LoraSLReward.reconstruct(...) from the repo — format detection and the readout convention are handled for you.

Training

LoRA (r16, lr 7e-5, 5-epoch schedule — best checkpoint is epoch 1; like the AV, it converges in one epoch) on ~158k (description, activation) pairs (_tokenpred_gpt4o_clean descriptions × 36 layers; 528 texts held out, ids ship as val_text_ids.json). Loss: direction MSE + InfoNCE contrastive (weight 1.0), mse_scale 59.87.

Evaluation

Validation cosine (held-out texts, raw): 0.942 overall. Per-layer:

layer band	val cosine
L0–L9 (early)	0.96–0.99
L10–L19 (mid)	0.95–0.96
L20–L29 (late)	0.90–0.94
L30–L34 (near-output)	0.91–0.95
L35 (final)	0.79

Full per-layer table in nla_meta.yaml. Round-trip numbers (verbalizer → this AR → centered cosine on a clean 284-text holdout) live on the AV and AV-GRPO cards.

Limitations

Reconstruction quality tracks description quality — it cannot conjure geometry a vague description doesn't carry (that asymmetry is the point). Raw cosine at these layers is dominated by the shared per-layer mean; always compare centered cosines (per-layer mean removed) when auditing descriptions. Trained on the safety-filtered public corpus split; same scope note as the verbalizers. Qwen3 chat-template use elsewhere in the pipeline needs enable_thinking=False (the AR prompt itself is raw, no template).