Instructions to use playforgecoding/LFM2-1.2B-Extract-facts with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use playforgecoding/LFM2-1.2B-Extract-facts with Transformers:

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

pipe = pipeline("text-generation", model="playforgecoding/LFM2-1.2B-Extract-facts")
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("playforgecoding/LFM2-1.2B-Extract-facts")
model = AutoModelForCausalLM.from_pretrained("playforgecoding/LFM2-1.2B-Extract-facts", 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 playforgecoding/LFM2-1.2B-Extract-facts with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "playforgecoding/LFM2-1.2B-Extract-facts"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "playforgecoding/LFM2-1.2B-Extract-facts",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/playforgecoding/LFM2-1.2B-Extract-facts

SGLang

How to use playforgecoding/LFM2-1.2B-Extract-facts 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 "playforgecoding/LFM2-1.2B-Extract-facts" \
    --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": "playforgecoding/LFM2-1.2B-Extract-facts",
		"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 "playforgecoding/LFM2-1.2B-Extract-facts" \
        --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": "playforgecoding/LFM2-1.2B-Extract-facts",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use playforgecoding/LFM2-1.2B-Extract-facts with Docker Model Runner:
```
docker model run hf.co/playforgecoding/LFM2-1.2B-Extract-facts
```

LFM2-1.2B-Extract-facts / README.md

playforgecoding

Model card

cc29429 verified 14 days ago

preview code

Raw History Blame Contribute Delete

4.23 kB

---
license: other
license_name: lfm1.0
license_link: LICENSE
language:
  - en
pipeline_tag: text-generation
library_name: transformers
base_model: LiquidAI/LFM2-1.2B-Extract
datasets:
  - playforgecoding/spelling-creator-fact-claims
tags:
  - liquid
  - lfm2
  - edge
  - extraction
  - fact-checking
  - spelling-creator
  - trl
  - sft
---
# LFM2-1.2B-Extract, fine-tuned to list checkable facts in lesson passages

The merged weights (bf16, safetensors). For the browser or Node, use the ONNX
export at
[playforgecoding/LFM2-1.2B-Extract-facts-ONNX](https://huggingface.co/playforgecoding/LFM2-1.2B-Extract-facts-ONNX).

**Experimental.** It is not used by Spelling Creator: on held-out lessons it
still lists a few facts the rules leave out, and two of them turned into false
alarms in a fact check (see Results).

## What it does

Reads one passage of a lesson and lists the facts in it that can be checked
against Wikidata: a number, a date or a named thing stated about one specific,
named, real thing ("Mount Everest rises 8,849 METRES", "CANBERRA is the
capital"). Each claim has the exact words it rests on (`quote`), what it is
about (`subject`, `kind`), which `property` it states, and the `value` and
`unit` or the `stated` name. It never says whether a fact is right; that is
done in code against Wikidata.

It is a fine-tune of
[LiquidAI/LFM2-1.2B-Extract](https://huggingface.co/LiquidAI/LFM2-1.2B-Extract)
for [Spelling Creator](https://github.com/Spelling-Creator/spelling-creator), a
lesson builder for Spelling to Communicate (S2C), where lessons are read aloud
to nonspeaking spellers. Spelling Creator's fact check asks a hosted model for
the claims today; a model like this one would let the whole check run on the
device.

## Training

LoRA (rank 16, learning rate 2e-4, 5 epochs, 3072-token context) on the
[spelling-creator-fact-claims](https://huggingface.co/datasets/playforgecoding/spelling-creator-fact-claims)
dataset: 347 passages and 750 claims, from 10 hub lessons and 116 sections
written for coverage, 96 of the passages with nothing to list. The labels are
Claude Opus's, made with the same prompt Spelling Creator's server sends its
own provider. Only the reply carries loss. Two hub lessons (24 passages, 20
claims) are held out. The scripts are in the Spelling Creator repo under
`packages/core/scripts/fact-eval/`.

## Results

The int8 ONNX export on the 24 held-out passages, beside the same training on
the 350M model:

| metric                                        |       350M | this model |
| --------------------------------------------- | ---------: | ---------: |
| replies in the schema                         |      24/24 |      24/24 |
| labelled claims found                         |        50% |        55% |
| listed claims that match a label              |        42% |        46% |
| quotes not in the passage                     |          2 |          2 |
| empty list for a passage with nothing to list |       8/13 |      10/13 |
| subject named right, of the claims found      |        50% |        64% |
| value, unit and qualifier right, of those     | 90 to 100% |       100% |

Run through the real checker against Wikidata, its claims give 2 false alarms
(a right fact reported as contradicted by Wikidata) in those 24 passages, and
miss 7 of the 14 findings the labels give. Both false alarms are facts the
rules leave out: the year Pompeii was rediscovered, checked as when it began,
and how long sunlight takes to reach Earth, checked as Earth's orbital period.
About 10 seconds a passage on an M4's CPU. The write-up is at
[spellingcreator.org/docs/monorepo/fact-claim-model](https://spellingcreator.org/docs/monorepo/fact-claim-model).

## Prompt

The system prompt is the rules (which facts count, the properties with a short
description each, how to choose between near ones, what each field means) and
the JSON Schema of the reply; the user turn is `Lesson: <title>`, a blank line
and the passage. Greedy decoding. The exact text is `passageClaimsSystemPrompt`
in `packages/core/src/factClaims.js`, and every training example carries it.

## Licence

Derived from LFM2-1.2B-Extract and released under the same LFM Open License
v1.0 (see LICENSE).

	---
	license: other
	license_name: lfm1.0
	license_link: LICENSE
	language:
	- en
	pipeline_tag: text-generation
	library_name: transformers
	base_model: LiquidAI/LFM2-1.2B-Extract
	datasets:
	- playforgecoding/spelling-creator-fact-claims
	tags:
	- liquid
	- lfm2
	- edge
	- extraction
	- fact-checking
	- spelling-creator
	- trl
	- sft
	---

	# LFM2-1.2B-Extract, fine-tuned to list checkable facts in lesson passages

	The merged weights (bf16, safetensors). For the browser or Node, use the ONNX
	export at
	[playforgecoding/LFM2-1.2B-Extract-facts-ONNX](https://huggingface.co/playforgecoding/LFM2-1.2B-Extract-facts-ONNX).

	Experimental. It is not used by Spelling Creator: on held-out lessons it
	still lists a few facts the rules leave out, and two of them turned into false
	alarms in a fact check (see Results).

	## What it does

	Reads one passage of a lesson and lists the facts in it that can be checked
	against Wikidata: a number, a date or a named thing stated about one specific,
	named, real thing ("Mount Everest rises 8,849 METRES", "CANBERRA is the
	capital"). Each claim has the exact words it rests on (`quote`), what it is
	about (`subject`, `kind`), which `property` it states, and the `value` and
	`unit` or the `stated` name. It never says whether a fact is right; that is
	done in code against Wikidata.

	It is a fine-tune of
	[LiquidAI/LFM2-1.2B-Extract](https://huggingface.co/LiquidAI/LFM2-1.2B-Extract)
	for [Spelling Creator](https://github.com/Spelling-Creator/spelling-creator), a
	lesson builder for Spelling to Communicate (S2C), where lessons are read aloud
	to nonspeaking spellers. Spelling Creator's fact check asks a hosted model for
	the claims today; a model like this one would let the whole check run on the
	device.

	## Training

	LoRA (rank 16, learning rate 2e-4, 5 epochs, 3072-token context) on the
	[spelling-creator-fact-claims](https://huggingface.co/datasets/playforgecoding/spelling-creator-fact-claims)
	dataset: 347 passages and 750 claims, from 10 hub lessons and 116 sections
	written for coverage, 96 of the passages with nothing to list. The labels are
	Claude Opus's, made with the same prompt Spelling Creator's server sends its
	own provider. Only the reply carries loss. Two hub lessons (24 passages, 20
	claims) are held out. The scripts are in the Spelling Creator repo under
	`packages/core/scripts/fact-eval/`.

	## Results

	The int8 ONNX export on the 24 held-out passages, beside the same training on
	the 350M model:

	\| metric \| 350M \| this model \|
	\| --------------------------------------------- \| ---------: \| ---------: \|
	\| replies in the schema \| 24/24 \| 24/24 \|
	\| labelled claims found \| 50% \| 55% \|
	\| listed claims that match a label \| 42% \| 46% \|
	\| quotes not in the passage \| 2 \| 2 \|
	\| empty list for a passage with nothing to list \| 8/13 \| 10/13 \|
	\| subject named right, of the claims found \| 50% \| 64% \|
	\| value, unit and qualifier right, of those \| 90 to 100% \| 100% \|

	Run through the real checker against Wikidata, its claims give 2 false alarms
	(a right fact reported as contradicted by Wikidata) in those 24 passages, and
	miss 7 of the 14 findings the labels give. Both false alarms are facts the
	rules leave out: the year Pompeii was rediscovered, checked as when it began,
	and how long sunlight takes to reach Earth, checked as Earth's orbital period.
	About 10 seconds a passage on an M4's CPU. The write-up is at
	[spellingcreator.org/docs/monorepo/fact-claim-model](https://spellingcreator.org/docs/monorepo/fact-claim-model).

	## Prompt

	The system prompt is the rules (which facts count, the properties with a short
	description each, how to choose between near ones, what each field means) and
	the JSON Schema of the reply; the user turn is `Lesson: <title>`, a blank line
	and the passage. Greedy decoding. The exact text is `passageClaimsSystemPrompt`
	in `packages/core/src/factClaims.js`, and every training example carries it.

	## Licence

	Derived from LFM2-1.2B-Extract and released under the same LFM Open License
	v1.0 (see LICENSE).