Instructions to use Prience91/Climate-1000-Qwen-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Prience91/Climate-1000-Qwen-4b with Transformers:

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

pipe = pipeline("text-generation", model="Prience91/Climate-1000-Qwen-4b")
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("Prience91/Climate-1000-Qwen-4b")
model = AutoModelForCausalLM.from_pretrained("Prience91/Climate-1000-Qwen-4b", 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 Prience91/Climate-1000-Qwen-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Prience91/Climate-1000-Qwen-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Prience91/Climate-1000-Qwen-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Prience91/Climate-1000-Qwen-4b

SGLang

How to use Prience91/Climate-1000-Qwen-4b 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 "Prience91/Climate-1000-Qwen-4b" \
    --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": "Prience91/Climate-1000-Qwen-4b",
		"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 "Prience91/Climate-1000-Qwen-4b" \
        --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": "Prience91/Climate-1000-Qwen-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Prience91/Climate-1000-Qwen-4b with Docker Model Runner:
```
docker model run hf.co/Prience91/Climate-1000-Qwen-4b
```

GIIS Mini Qwen3.5-4B merged BF16 model

Standalone text-generation model formed by merging the selected checkpoint-230 LoRA adapter into the original unquantized BF16 Qwen3.5-4B base. The base model weights are included; a separate adapter load is not required.

Provenance

Base: Qwen/Qwen3.5-4B, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
Source adapter: Prience91/GIIS-Mini-Qwen3.5-4B-LoRA-adapter, commit 2c3a529c35218e07ad8443fcb57cb5e877f57a3e.
Selected checkpoint 230 (end of epoch 1); validation loss 0.37746524810791016.
Full SFT completed 460 steps and two epochs on 22 October 2026.
Final step-460 validation loss was 0.38062456250190735.
Training used a frozen BF16 base, BF16 mixed precision, and FP32 LoRA adapters; rank 32, alpha 64, dropout 0.05, supported linear layers, completion-only loss.
Same V17-Turbo release as the Qwen3.8-27B experiment: 3,665 training examples, 406 validation examples; learning rate 1e-4, microbatch 1, accumulation 16, 8,192-token ceiling, seed 3407, and thinking disabled.
The stopped 4B run was resumed from checkpoint 120 with periodic CUDA cache clearing and telemetry, preserving optimization settings and the dataset.

Usage

Load with transformers.AutoModelForCausalLM and AutoTokenizer. Architecture: Qwen3_5ForCausalLM, model type qwen3_5_text; weights BF16, no quantization. Use a compatible Transformers version (merge validated with 5.15.1), and apply the included chat template with enable_thinking=False to match training. This export contains the trained text model, not the unused vision encoder.

Validation and limitations

All adapter tensors were checked against the source before merging. Safe merge, finite-weight checks, exact merge arithmetic in layers 0/15/31, standalone model reload, and tokenizer/chat-template loading passed. See merge_manifest.json.

The reported losses come from adapter training; they were not remeasured on the merged export. No generation-quality evaluation is claimed for this export. The heldout benchmark remains provisional (98 authored/reviewed cases, 78 approved); token loss does not establish scientific accuracy or citation quality.