neural-txt 0.1.4

Structured NLP tasks powered by a fine-tuned small language model

neural-txt

Structured NLP tasks powered by a fine-tuned 135M parameter language model. Extract bullets, generate Q&A pairs, build knowledge graphs, and more — all running locally. Narrow vertical local intelligence that runs super cheaply in resource constrained envs.

https://github.com/user-attachments/assets/04774af0-dc51-42e7-b2a6-d6f50bf4e258

Support

If you find this helpful, consider supporting on Patreon — it hosts all code, projects, slides, and write-ups from the YouTube channel.

Install

# Base (no inference backend)
pip install neural-txt

# With HuggingFace backend (torch)
pip install neural-txt[hf]

# With MLX backend (Apple Silicon)
pip install neural-txt[mlx]

NeuralTxtReward works with either backend: install neural-txt[hf] for the Hugging Face / torch scorer, or neural-txt[mlx] for Apple Silicon MLX.

Quick start

from neuraltxt import NeuralTxt

model = NeuralTxt(backend="mlx")  # or backend="hf"

passage = """
Transformers have revolutionized NLP by introducing the self-attention
mechanism. Unlike RNNs, transformers process all tokens in parallel,
leading to significant training speedups.
"""

# Extract key points
bullets = model.extract_bullets(passage)

# Generate question-answer pairs
pairs = model.generate_qa_pairs(passage)

# Extract knowledge graph triplets
triplets = model.extract_triplets(passage)

Reward scoring

NeuralTxtReward scores generated responses against a reference answer with paperbd/neuraltxt-reward-22M. Use it to score one answer, score a batch, or rank candidate responses.

from neuraltxt import NeuralTxtReward

reward = NeuralTxtReward(backend="mlx")  # or backend="hf"

reference = "The capital of France is Paris."
responses = [
    "Paris is the capital of France.",
    "France's capital is Lyon.",
]

score = reward.score(responses[0], reference)          # float between 0 and 1
scores = reward.batch_score(responses, reference)      # list[float]
ranked = reward.rank(responses, reference)             # list[RankedResponse]

for item in ranked:
    print(item.index, item.score, item.response)

rank() preserves the original response index and sorts highest score first. Pass a local model directory with NeuralTxtReward("path/to/reward-model").

Beam candidates

Generation methods accept num_beams with a default of 1. The public methods still return one parsed result: the first / highest-ranked candidate. With the HuggingFace backend, num_beams is forwarded as beam search with num_return_sequences=num_beams. With MLX, candidates are generated the same way as the existing repeated generation path.

bullets = model.extract_bullets(passage, num_beams=4)

See examples/beam_candidates.py for a complete example, including how to inspect all raw beam candidates.

JSON mode

Every method supports json=True for guaranteed structured output via outlines:

# Returns a BulletsOutput pydantic model
bullets = model.extract_bullets(passage, json=True)
print(bullets.bullets)  # list[str]

# Returns a QAPairsOutput pydantic model
qa = model.generate_qa_pairs(passage, json=True)
for pair in qa.pairs:
    print(pair.question, pair.answer)

# Returns a TripletsOutput pydantic model
triplets = model.extract_triplets(passage, json=True)
for t in triplets.triplets:
    print(t.subject, t.relation, t.object)

API

Generation API

Method	Input	Output	JSON Output
extract_bullets(passage)	passage	list[str]	BulletsOutput
generate_qa_pairs(passage)	passage	list[QAPair]	QAPairsOutput
generate_question(passage)	passage	str	QuestionOutput
generate_questions_list(passage)	passage	list[str]	QuestionsListOutput
extract_fact(passage)	passage	str	FactOutput
answer(question, passage)	question + passage	str	AnswerOutput
rephrase(passage)	passage	str	RephraseOutput
continue_from(passage)	passage start	str	ContinuationOutput
extract_triplets(passage)	passage	list[Triplet]	TripletsOutput
compare(passage_a, passage_b)	two passages	str	ComparisonOutput
find_relevant(question, passages)	question + passage list	RetrievalResult	RetrievalOutput

Reward API

Method	Input	Output
score(response, reference)	one response + reference answer	float
batch_score(responses, reference)	response list + reference answer	list[float]
rank(responses, reference)	response list + reference answer	list[RankedResponse]

NeuralTxtReward accepts backend="hf" or backend="mlx".

Models

Interface	Default model
NeuralTxt(backend="hf")	paperbd/neuraltxt-v1-135M
NeuralTxt(backend="mlx")	paperbd/neuraltxt-v1-135M-mlx
NeuralTxtReward(backend="hf")	paperbd/neuraltxt-reward-22M
NeuralTxtReward(backend="mlx")	paperbd/neuraltxt-reward-22M-mlx

Pass a custom path: NeuralTxt("path/to/model", backend="hf")

• Training dataset: paperbd/paper_instructions_300K-v1
• Synthetic data generation: text-albumentations

Gradio demo

pip install neural-txt[app]

# HuggingFace (default)
python app.py

# MLX (Apple Silicon)
python app.py --mlx

# Options
#   --temperature 0.4    sampling temperature (default 0.4)
#   --num-beams 2        beam candidates, 1-4 (default 1)