Instructions to use Lexsi/tiny-aya-global-hh-rlhf-aligned with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Lexsi/tiny-aya-global-hh-rlhf-aligned with Transformers:

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

pipe = pipeline("text-generation", model="Lexsi/tiny-aya-global-hh-rlhf-aligned")
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("Lexsi/tiny-aya-global-hh-rlhf-aligned")
model = AutoModelForCausalLM.from_pretrained("Lexsi/tiny-aya-global-hh-rlhf-aligned", 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 Lexsi/tiny-aya-global-hh-rlhf-aligned with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Lexsi/tiny-aya-global-hh-rlhf-aligned"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Lexsi/tiny-aya-global-hh-rlhf-aligned",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Lexsi/tiny-aya-global-hh-rlhf-aligned

SGLang

How to use Lexsi/tiny-aya-global-hh-rlhf-aligned 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 "Lexsi/tiny-aya-global-hh-rlhf-aligned" \
    --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": "Lexsi/tiny-aya-global-hh-rlhf-aligned",
		"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 "Lexsi/tiny-aya-global-hh-rlhf-aligned" \
        --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": "Lexsi/tiny-aya-global-hh-rlhf-aligned",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Lexsi/tiny-aya-global-hh-rlhf-aligned with Docker Model Runner:
```
docker model run hf.co/Lexsi/tiny-aya-global-hh-rlhf-aligned
```

Tiny-Aya-Global: HH-RLHF Safety Alignment

CohereLabs/tiny-aya-global safety-aligned with DPO on Anthropic HH-RLHF (harmless-base), using the SafeTune paper's parameters.

What we did

We aligned CohereLabs/tiny-aya-global for safety with Direct Preference Optimization (LoRA, merged into the base weights) on the harmless-base split of Anthropic/hh-rlhf. You get one drop-in checkpoint: the LoRA adapter is already merged, so it is loaded and served exactly like the base model.

Built using SafeTune, Lexsi's library for auditing and correcting safety drift.

Training parameter	Value
Method	DPO + LoRA (TRL 0.23.0), adapter merged into bf16 base
Dataset	Anthropic/hh-rlhf, `harmless-base`, train, 20,000 samples
β	0.1
Learning rate	2e-5 (cosine, warmup ratio 0.03)
Epochs	1 (1,249 steps)
Batch size	2 per device × 8 gradient accumulation
LoRA	r = 16, α = 32, dropout 0.05; q, k, v, o, gate, up, down projections
Max length	1,024 tokens (prompt 512)
Precision	bf16


Base model	`CohereLabs/tiny-aya-global`
Developed by	Lexsi Labs (Lithasa Technologies Pvt. Ltd.)
Model size	3.35B parameters
Context length	8K tokens (inherited from the base model)
Languages	70+ (inherited from the base model)
License	CC-BY-NC 4.0 (inherited from the base model); see License
Contact	support@lexsi.ai

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Lexsi/tiny-aya-global-hh-rlhf-aligned"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "How can I stay safe when meeting someone from an online forum?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)

out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

License

This is a derivative work of CohereLabs/tiny-aya-global, which Cohere Labs releases under CC-BY-NC 4.0 with its Acceptable Use Policy. This checkpoint inherits those terms: non-commercial use with attribution. Please also follow the Cohere Labs Acceptable Use Policy.

Contact

Questions or errors in this card: support@lexsi.ai

Downloads last month: 5

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for Lexsi/tiny-aya-global-hh-rlhf-aligned

Base model

CohereLabs/tiny-aya-base

Finetuned

CohereLabs/tiny-aya-global

Finetuned

(60)

this model