qwen3-4b-argument-analyst-r6

Qwen3-4B Argument Analyst — r6 Merged

This is a merged version of the deployed r6 argument-analysis LoRA adapter and Qwen/Qwen3-4B. It is a research prototype for turning English prose arguments into a constrained JSON decomposition.

Intended behavior

The model is prompted to emit:

{
  "claim": "the conclusion",
  "warrants": ["each distinct inferential step"],
  "impact": "the stake or consequence",
  "implicit_premises": ["minimally necessary unstated assumptions"],
  "load_bearing_component": "one listed component, or an empty abstention",
  "negation_justification": "why negating that component breaks the argument",
  "support_edges": [["w0", "claim"]]
}

The Part-2 support graph uses claim, w{k}, i{k}, and impact component tokens.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "fzhualpha/qwen3-4b-argument-analyst-r6"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    torch_dtype=torch.bfloat16,
    device_map="auto",
).eval()

system = """You are an argument analyst. Given a prose argument, return only one valid JSON object.
Faithfully identify the claim, each warrant step, impact, and only minimally necessary implicit premises.
Also emit support_edges using claim, w{k}, i{k}, and impact tokens. Use an empty load-bearing component
when no unique component exists; never choose a component merely because it sounds weak."""

argument = "The city should add bus lanes because buses move more people in less road space."
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": argument},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=768,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )
print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Training

• Base: Qwen/Qwen3-4B
• Method: bf16 LoRA supervised fine-tuning
• LoRA rank/alpha/dropout: 16 / 32 / 0.05
• Target modules: attention projections and MLP projections
• Objective: completion-only JSON generation
• Associated dataset configuration: fzhualpha/argument-analyst-data, configuration r6

The original run did not preserve a complete immutable manifest, so exact package versions and the precise adapter hash are not claimed here.

Evaluation summary

The most defensible current result is a narrow reliability trade-off on a train-disjoint 90-row args.me set:

• Seven core deterministic checks: 90.0% for the tuned model versus 72.2% for Claude Opus.
• Three-model content-judge pass: 13.3% for the tuned model versus 52.2% for Claude Opus.

The model therefore learned strong schema and extraction discipline, but it remains substantially weaker on faithful real-prose semantic decomposition. Its r6 load-bearing result was measured on a repeatedly used development set, not an untouched final test. The model also passed 9 of 12 canonical golden cases.

Limitations

• This is not demonstrated to be a generally superior argument analyst.
• The deterministic support-graph solver checks consistency with a model-authored graph; it does not independently establish that the graph faithfully represents the prose.
• Empty or incomplete structures can exploit weaknesses in the current mechanical checks.
• Implicit-premise and no-stake cases remain weak.
• Paraphrase invariance has not been adequately measured for this model.
• Generation is not grammar-constrained, so invalid JSON remains possible.
• The training data is predominantly synthetic and contains incomplete provenance.
• No final evaluation set is both family-disjoint and untouched after model selection.

Use human review for consequential analysis and independently validate every support graph or load-bearing conclusion.