Instructions to use AxisMeru/nyaya-element-judge-4b-v6 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AxisMeru/nyaya-element-judge-4b-v6 with Transformers:

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

pipe = pipeline("text-generation", model="AxisMeru/nyaya-element-judge-4b-v6")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AxisMeru/nyaya-element-judge-4b-v6")
model = AutoModelForCausalLM.from_pretrained("AxisMeru/nyaya-element-judge-4b-v6", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AxisMeru/nyaya-element-judge-4b-v6 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AxisMeru/nyaya-element-judge-4b-v6"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxisMeru/nyaya-element-judge-4b-v6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AxisMeru/nyaya-element-judge-4b-v6

SGLang

How to use AxisMeru/nyaya-element-judge-4b-v6 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 "AxisMeru/nyaya-element-judge-4b-v6" \
    --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": "AxisMeru/nyaya-element-judge-4b-v6",
		"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 "AxisMeru/nyaya-element-judge-4b-v6" \
        --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": "AxisMeru/nyaya-element-judge-4b-v6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AxisMeru/nyaya-element-judge-4b-v6 with Docker Model Runner:
```
docker model run hf.co/AxisMeru/nyaya-element-judge-4b-v6
```

nyaya-element-judge-4b-v6 (full fine-tune)

Part of Prabhāsa-Nyāya, our own trained Indian-legal models behind the Pravrudhi-Nyaya product. This version is not the one the product serves.

Status: Not promoted. Not for use. This model was trained but not promoted, and it does not replace the production judges. A check of its training data found text that is also in the training split of a non-commercial benchmark. That overlap has not been cleared. This model is published for inspection and kept here as a record only. The training rows are not published. It is not a released or promoted model, and no performance claim is made on this card.

Label: v6. A full fine-tune continuing the earlier 4B element judge.

Training data: real court-derived rows plus constructed rows. The real rows were extracted from court judgments by a pipeline with two model readers and no human review; the labels are therefore pipeline-measured, not human-reviewed.

Some of the not-established training rows carry a generic or weak supporting quote. Every label comes from two model readers that agreed; no human reviewed them. The checker that admits training rows is being changed to test that a quote speaks to its element; this model was trained before that change.

In plain words: a model could learn that rows from one source tend to carry one label, and this was not controlled.

Known uncontrolled shortcut channel: source-correlated label rate (gate A2 T2 not binding for mixed-source TRAINs per D-GATE-A2-MIXED rev2); reported descriptively; safety rests on the post-training stops.

Licence and attribution

Licence: Apache License 2.0, the licence of the base model Qwen/Qwen3-4B (its card and LICENSE read on 17 October 2026). This is a modified (fully fine-tuned) copy of the base model and is released under the same licence; see LICENSE and NOTICE.

Credit: some of the training examples are built from the Indian High Court Judgments data set (Dattam Labs, licensed CC BY 4.0, https://registry.opendata.aws/indian-high-court-judgments), selected and annotated by us.

Files: LICENSE, NOTICE, model.safetensors, config.json, generation_config.json, tokenizer files, chat_template.jinja, training_args.bin, FINAL_SHAS.txt (per-file sha256 of the trainer's output; verify with sha256sum -c FINAL_SHAS.txt).

Training setup (recipe): full fine-tune, bf16, 8-bit AdamW, learning rate 2e-6 cosine, effective batch 16 (micro-batch 2 x accumulation 8), max length 2048, one epoch, seed 42, one run, final step only. Trained on a single H100 80GB pod (not the 5090).

Dev-stack, local; exploratory; no promotion and no improvement claim.

Not legal advice; a lawyer decides.

Downloads last month: 895

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for AxisMeru/nyaya-element-judge-4b-v6

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Finetuned

(1202)

this model