Instructions to use Eclipse-Senpai/KeyLM-75M-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Eclipse-Senpai/KeyLM-75M-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="Eclipse-Senpai/KeyLM-75M-Instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Eclipse-Senpai/KeyLM-75M-Instruct", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Eclipse-Senpai/KeyLM-75M-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Eclipse-Senpai/KeyLM-75M-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Eclipse-Senpai/KeyLM-75M-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Eclipse-Senpai/KeyLM-75M-Instruct

SGLang

How to use Eclipse-Senpai/KeyLM-75M-Instruct 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 "Eclipse-Senpai/KeyLM-75M-Instruct" \
    --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": "Eclipse-Senpai/KeyLM-75M-Instruct",
		"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 "Eclipse-Senpai/KeyLM-75M-Instruct" \
        --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": "Eclipse-Senpai/KeyLM-75M-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Eclipse-Senpai/KeyLM-75M-Instruct with Docker Model Runner:
```
docker model run hf.co/Eclipse-Senpai/KeyLM-75M-Instruct
```

Eclipse-Senpai commited on 3 days ago

Commit

b433cbb

verified ·

1 Parent(s): f15c3a0

Update knowledge/reasoning table (transformers harness), base+instruct consistent

Browse files

Files changed (1) hide show

README.md +4 -4

README.md CHANGED Viewed

@@ -103,15 +103,15 @@ KeyLM beats the original SmolLM-135M-Instruct at roughly half the size and a fra
 ### Knowledge and reasoning
-On standard multiple-choice benchmarks KeyLM performs at or near random chance. This is the expected trade-off at 75M parameters and 18B tokens: the model holds little parametric knowledge, and instruction tuning changes its behavior, not its knowledge. Scores are zero-shot via `lm_eval` (accuracy; ARC and HellaSwag use length-normalized accuracy).
 | Model | MMLU | ARC (avg) | HellaSwag | PIQA | WinoGrande | OpenBookQA |
 |---|---|---|---|---|---|---|
-| KeyLM-75M (base) | 23.0 | 26.4 | — | 52.9 | 48.3 | 19.8 |
-| **KeyLM-75M-Instruct** | **23.0** | **26.1** | **26.7** | **53.1** | **48.9** | **18.4** |
 | Random baseline | 25.0 | 25.0 | 25.0 | 50.0 | 50.0 | 25.0 |
-Instruction tuning leaves knowledge and reasoning essentially unchanged: the base and instruct checkpoints track each other and both sit close to the random baseline. The base model's HellaSwag score will be added with its release.
 ## Training

 ### Knowledge and reasoning
+On zero-shot multiple-choice benchmarks (`lm_eval`; accuracy, with length-normalized accuracy for ARC and HellaSwag) KeyLM is modest but above random on basic commonsense, and at chance on knowledge-heavy tasks. This is expected at 75M parameters and 18B tokens.
 | Model | MMLU | ARC (avg) | HellaSwag | PIQA | WinoGrande | OpenBookQA |
 |---|---|---|---|---|---|---|
+| KeyLM-75M (base) | 23.0 | 29.9 | 29.7 | 60.0 | 48.4 | 25.0 |
+| **KeyLM-75M-Instruct** | **24.0** | **30.8** | **31.0** | **61.3** | **48.3** | **25.0** |
 | Random baseline | 25.0 | 25.0 | 25.0 | 50.0 | 50.0 | 25.0 |
+Base and instruct track each other closely, so instruction tuning leaves knowledge and reasoning roughly unchanged. PIQA and ARC-easy land clearly above chance, while MMLU sits at the random baseline.
 ## Training