Instructions to use zack-r-s/hackring-LFM2.5-230M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use zack-r-s/hackring-LFM2.5-230M with Transformers:

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

pipe = pipeline("text-generation", model="zack-r-s/hackring-LFM2.5-230M")
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("zack-r-s/hackring-LFM2.5-230M")
model = AutoModelForCausalLM.from_pretrained("zack-r-s/hackring-LFM2.5-230M", 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 zack-r-s/hackring-LFM2.5-230M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "zack-r-s/hackring-LFM2.5-230M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zack-r-s/hackring-LFM2.5-230M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/zack-r-s/hackring-LFM2.5-230M

SGLang

How to use zack-r-s/hackring-LFM2.5-230M 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 "zack-r-s/hackring-LFM2.5-230M" \
    --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": "zack-r-s/hackring-LFM2.5-230M",
		"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 "zack-r-s/hackring-LFM2.5-230M" \
        --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": "zack-r-s/hackring-LFM2.5-230M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use zack-r-s/hackring-LFM2.5-230M with Docker Model Runner:
```
docker model run hf.co/zack-r-s/hackring-LFM2.5-230M
```

Hackring LFM2.5-230M: dataset/model v004

This merged full-weight checkpoint was fine-tuned from LiquidAI/LFM2.5-230M on the matching dataset v004. It does not continue the v002 or v003 checkpoints. The original LFM Open License v1.0 is included in LICENSE; use and redistribution remain subject to its terms. The shared repository's v002, v003, and v004 tags preserve their respective checkpoints; main serves v004.

What changed

The earlier checkpoints were trained to emit numerical channel_write PV setpoints. V004 is trained to convert an operator orbit-centering request into orbit_correct(plane, effort), where plane is X, Y, or both and effort is gentle, balanced, or aggressive. A separate deterministic bounded least-squares skill reads the fixed 24-PV simulated PyAT ring, computes corrector settings from the versioned response matrix, writes them, and measures the after-state. The model does not calculate kick values. The operator harness normalizes an unspecified effort to balanced and rejects a wrong explicit effort or wrong plane before writing. No Osprey or lattice configuration was changed for this pass.

The 432/72/72 train/validation/test conversations come from fresh seed-disjoint local PyAT disturbances; no GPT annotation was used. The Liquid Finetune run used assistant-only SFT, default rank-8 LoRA, two epochs and 108 optimizer steps on Apple M4 Pro MPS. Validation loss was 0.0381 then 0.0315. training_job.yaml, dataset_manifest.json, and training_results.json record the configuration and hashes. The source and version chart are in integration PR #8; the checkpoint's training commit remains on zrs-lfm-training.

Measured evaluation

Mark III is the complete PR #8 shared-fixture rerun. Its results index validates pinned model revisions, immutable case/task hashes, all seven full NeMo Gym runs (28 tasks × four repeats, zero rollout failures), and the separate local PyAT reports. The v001–v004 data were revalidated without regeneration. Mark I and the partial Mark II reports remain historical; the Mark II v004 Gym direct-write run used the wrong tool interface and is not a skill score.

Measure	Base 230M	Base 1.2B Instruct	v004
Local greedy CPU skill: exact plane and effort with measured improvement /28	0	0	12
Local balanced-default replay: measured improvement /28	14	—	28
Gym skill: measured improvement /112	73	28	72
Gym skill: raw plane-and-effort selection /112	31	0	42

The local greedy replay shows a v004 routing improvement on these prompts, but the corrected stochastic Gym skill benchmark does not show a deployed control-success advantage over base 230M (72/112 versus 73/112). A paired bootstrap over 28 tasks estimates the v004-minus-base Gym success difference at −0.9 percentage points, with a 95% interval of −22.3 to +20.5 points. Raw intent selection is +9.8 points with a −10.7 to +29.5 point interval. These fixed-task descriptive intervals do not establish broader generalization. The 1.2B model made many extra tool attempts, which the one-attempt rule rejects.

The shared-fixture API smoke passed BPM reporting, an exact setpoint request, an orbit-correction request, and an over-limit refusal. The browser smoke verified that chat, the Agent & goal panel, displayed PV, and reset reflect one PyAT fixture. Exact numeric writes and over-limit refusals are bounded harness behavior; v004 chooses the orbit-correction skill, whose deterministic solver supplies numerical PV writes. These are simulated-plant results, not real accelerator evidence or model-only inverse-optics ability.

V002 and v003 target direct numerical writes and are compared on that separate interface in the training/version chart. Their local 56-case scores were 24 and 26, but Gym direct-write success was 12/112 and 6/112; this does not establish v003 as the better controller. Local direct uses greedy Transformers MPS inference and a forced write-tool name; Gym uses vLLM at temperature 0.2 with one-attempt enforcement. Cross-interface or cross-backend counts should not be treated as one score.

Serve

vllm serve zack-r-s/hackring-LFM2.5-230M --revision v004 \
  --served-model-name LFM2.5-230M --enable-auto-tool-choice \
  --tool-call-parser lfm2 --host 127.0.0.1 --port 8000

The lfm2 tool parser is required. The checked-in Hackring chat agent supplies the skill implementation and checks the operator intent; serving this checkpoint alone does not provide accelerator control.

Downloads last month: 84

Safetensors

Model size

0.2B params

Tensor type

BF16

Model tree for zack-r-s/hackring-LFM2.5-230M

Base model

LiquidAI/LFM2.5-230M-Base

Finetuned

LiquidAI/LFM2.5-230M

Finetuned

(49)

this model

zack-r-s
/

hackring-LFM2.5-230M

Hackring LFM2.5-230M: dataset/model v004

What changed

Measured evaluation

Serve

Model tree for zack-r-s/hackring-LFM2.5-230M

Dataset used to train zack-r-s/hackring-LFM2.5-230M