Extract the logical skeleton of natural-language arguments into structured JSON.

nyashastino/lfx is a fine-tuned 3B parameter model (based on Qwen2.5-3B-Instruct) engineered specifically to deconstruct natural-language arguments. Given an argumentative passage, it extracts individual premises, isolates the asserted conclusion, determines whether the argument is deductive or inductive, identifies the specific logical form or fallacy (from a 22-class taxonomy), and reconstructs unstated enthymematic premises.

---

Quick Start

Run directly in your terminal:

ollama run nyashastino/lfx "All men are mortal. Socrates is a man. Therefore Socrates is mortal."

Output:

{
  "premises": [
    "All men are mortal.",
    "Socrates is a man."
  ],
  "conclusion": "Socrates is mortal.",
  "argument_type": "deductive",
  "form": [
    "categorical syllogism"
  ],
  "suppressed_premise": null
}

---

Available Tags

Tag	Quantization	Size	RAM Required	Best For
nyashastino/lfx:latest nyashastino/lfx:3b-q8	Q8_0 (8-bit)	3.3 GB	~4.5 GB	Primary Release. Best overall accuracy (89.7% benchmark accuracy, 78.1% on held-out human prose) and highest premise extraction fidelity.

To pull the model:

ollama pull nyashastino/lfx

---

Output Schema

The model responds strictly with a single, valid JSON object matching this schema:

{
  "premises": [
    "First premise statement",
    "Second premise statement"
  ],
  "conclusion": "The claim being argued for",
  "argument_type": "deductive" | "inductive",
  "form": [
    "primary logical form or fallacy"
  ],
  "suppressed_premise": "Unstated assumption required for validity, or null"
}

Field Definitions:

• premises (array of strings): Explicit claims provided as evidence or support.
• conclusion (string): Central asserted claim the premises aim to establish.
• argument_type (string): "deductive" (presented as following with necessity) or "inductive" (presented as probable/supported).
• form (array of 1–3 strings): Primary logical form or fallacy exhibited. Single-step arguments have 1 form; multi-step chains order moves with the primary move first.
• suppressed_premise (string or null): Reconstructed unstated premise (enthymeme) necessary to bridge premises to conclusion. null if the argument is complete as stated.

---

Supported Logical Forms (22 Classes)

Aligned with formal logic and critical thinking curricula (Duke University Think Again framework):

• Valid Deductive (6): modus ponens, modus tollens, hypothetical syllogism, disjunctive syllogism, categorical syllogism, reductio ad absurdum
• Inductive (6): generalization, application of generalization, inference to the best explanation, analogy, causal, authority
• Formal Fallacies (2): affirming the consequent, denying the antecedent
• Informal Fallacies (7): ad hominem, hasty generalization, false dilemma, ad populum, begging the question, straw man, equivocation
• Other (1): other (reserved for multi-step arguments or patterns outside this taxonomy)

---

Benchmark Performance

Evaluated on the unified LFX-Bench v1 (1,521 unseen argumentative passages across all 22 classes with verified provenance):

Benchmark Metric	Full Benchmark (n=1521)	Held-Out Human Prose (n=485)
End-to-End Form Accuracy	89.7% (1365 / 1521)	78.1% (379 / 485)
Macro-F1 (22 Classes)	0.861	0.840
Schema Validity	99.8%	99.8%
JSON Parses	99.9%	100%
Argument Type Accuracy	94.4%	84.5%
Premise F1	0.890	0.748
Conclusion Similarity	0.911	0.925
Suppressed Premise Match	0.900	0.812

Breakdown by Provenance Tier:

• constructed-symbolic (n=693): 98.8% (label specified by construction, verified by blind formalization through symbolic rules)
• constructed-consensus (n=309): 92.9% (inductive forms verified by blind multi-model agreement)
• external-human (n=485): 78.1% (authentic human op-eds and speech passages from LOGIC)
• project-authored (n=34): 41.2% (condensed classical philosophical texts; 12 of 34 are multi-step other)

Per-Class Performance on Key Forms:

• hypothetical syllogism: 1.000 F1 (100% recall, 100% precision)
• affirming the consequent: 0.989 F1 (98.9% recall, 98.9% precision)
• reductio ad absurdum: 0.975 F1 (98.8% recall, 96.3% precision)
• categorical syllogism: 0.922 F1 (98.9% recall, 86.4% precision)
• ad hominem: 0.843 F1 (75.0% recall, 96.2% precision)
• ad populum: 0.831 F1 (82.7% recall, 83.5% precision)
• false dilemma: 0.800 F1 (83.0% recall, 77.2% precision)
• begging the question: 0.788 F1 (71.1% recall, 88.5% precision)
• straw man: 0.681 F1 (85.2% recall, 56.8% precision)
• equivocation: 0.643 F1 (64.3% recall, 64.3% precision)

---

Code Examples

Python (ollama-python)

import json
import ollama

client = ollama.Client()

prompt = (
    "Either humanity dramatically reduces carbon emissions within a decade, "
    "or we face ecological collapse. We will not cut emissions, therefore "
    "ecological collapse is inevitable."
)

res = client.chat(
    model="nyashastino/lfx",
    messages=[{"role": "user", "content": prompt}],
    options={"temperature": 0.0},
)

result = json.loads(res["message"]["content"])
print(json.dumps(result, indent=2))

REST API (curl)

curl http://localhost:11434/api/chat -d '{
  "model": "nyashastino/lfx",
  "messages": [
    {"role": "user", "content": "Nobody has proven that ghosts don'\''t exist, so they must be real."}
  ],
  "stream": false
}'

JavaScript / TypeScript

import ollama from 'ollama';

const response = await ollama.chat({
  model: 'nyashastino/lfx',
  messages: [{
    role: 'user',
    content: 'Smartphones cause attention deficits because constantly switching apps fragments concentration.'
  }],
});

const analysis = JSON.parse(response.message.content);
console.log(analysis);

---

Limitations

1. Extraction, Not Fact-Checking: LFX is an extraction parser, not an epistemic fact-checker. It extracts the internal structural relationship between premises and conclusions; it does not evaluate whether the premises are true in the real world.
2. Closed Taxonomy: Arguments based on informal fallacies outside the 22 categories (e.g., slippery slope, no true Scotsman) will be mapped to the closest applicable category or labeled other.
3. Non-Argument Text: When fed pure descriptive prose (e.g., a weather report or narrative text), the model will attempt to find an argument. Ensure input passages contain an intended point or argument.

---

License & Attribution

• Base Weights: Fine-tuned from Qwen2.5-3B-Instruct under the Qwen Research License Agreement (Alibaba Cloud). Non-commercial research use only.
• Attribution: Built with Qwen.
• Dataset Attribution: Incorporates the LOGIC benchmark (Jin et al., Findings of EMNLP 2022), licensed under MIT.
• Source Code & Data: https://github.com/NyashaEysenck/lfx