You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

TunedAI Causal Reasoning 7B (Preview)

A fine-tuned variant of Qwen 2.5-7B-Instruct trained to produce structured causal reasoning output across observation, mechanism, projection, and counterfactual tiers as a default behavior of the weights.

This model is gated and provided under a commercial evaluation license. Access is granted on request for integration evaluation and research collaboration.

Model Overview

Property	Value
Base model	Qwen 2.5-7B-Instruct
Adaptation method	LoRA (Low-Rank Adaptation)
LoRA rank	64
Target modules	Attention layers only (Q, K, V, O projections)
Trainable parameters	~40M
Context window	128K tokens (inherited from base)
Training framework	PEFT + transformers
Inference framework	transformers, vLLM, MLX (quantized variant available)

What this model does

The model produces structured causal reasoning output, tagged across four tiers, as default weight behavior — without a system prompt or external scaffolding:

T1 — Observation: what the data, records, or facts directly show. Correlations, patterns, documented outcomes.
T2 — Mechanism: the causal pathway. How and why the observed outcome occurred. Underlying forces, incentives, processes.
T3 — Projection: given the mechanism, expected forward effects. Probable outcomes, downstream consequences.
T4 — Simulation: counterfactual analysis. What would have happened under different conditions. Alternate scenarios.

Output is reliably tagged in a consistent format that downstream code can parse to surface specific reasoning tiers as discrete elements (e.g., flagging a counterfactual recommendation in a UI).

Intended use cases

Agent reasoning layers — for systems that need parseable structured causal analysis at specific decision points (relationship triage, calendar restructuring, priority ranking, root-cause analysis)
Decision support tooling — applications where the user benefits from seeing the reasoning structure, not just the conclusion
Domain-specific analysis — finance (causal attribution, fraud root-cause), healthcare (treatment effect estimation), legal (but-for causation), security (attack attribution), business operations (churn analysis, marketing ROI)
Research — comparative work on structured output fine-tuning vs. prompted/chained alternatives

Out-of-scope use

General-purpose chat (use base Qwen 2.5-7B-Instruct or larger general-purpose models)
Pure summarization or transcription tasks (specialized reasoning behavior is unnecessary overhead)
Tool-calling-heavy agentic flows without verification (see "Known limitations" below)
Safety-critical decisions without human oversight

Training data

Trained on a curated corpus of structured causal reasoning examples constructed under TunedAI Labs' methodology. Training data composition, generation procedure, and curriculum design are proprietary and not disclosed in this card. Available under NDA for partnership and licensing discussions.

Evaluation

See the accompanying eval card (EVAL_CARD.md) for full methodology and results. Headline numbers:

Benchmark	Base Qwen 2.5-7B	TunedAI Causal Reasoning 7B
CLadder (causal reasoning)	~63%	96.96%
Held-out evaluation (independent)	~63%	91.5%

Independent held-out evaluation conducted by Matt Wong, an outside applied AI engineer (~25 years in security and forensics, former White House).

Known limitations

Tool-calling adherence is not yet stress-tested under the LoRA. Base Qwen 2.5-7B supports OpenAI-compatible function calling natively. Behavior under heavy tool-call load with the causal-reasoning LoRA active has not been comprehensively evaluated. Treat tool-call performance as untested for production agentic flows pending forthcoming evaluation results.
Domain transfer. Training emphasized general causal reasoning. Behavior on specialized domains (medical, legal, financial) is plausible but not guaranteed; in-domain validation is recommended before production deployment.
Tier label consistency. Output is labeled in the vast majority of cases on in-distribution prompts, but failure modes exist (occasional missing tiers, particularly T3 on short-context prompts). Production parsers should handle the absence of an expected tier gracefully.
English-language only at this time. Multilingual variants are on the roadmap but not yet trained.
Not safety-tuned beyond the base model. Inherits Qwen 2.5's safety properties. Custom RLHF / safety fine-tuning has not been added.

Ethical considerations

Causal claims: the model produces causal-style output. Users should not interpret model output as ground-truth causal inference. The model identifies plausible causal structure based on training; verifying claims against domain knowledge and data remains the user's responsibility.
Human oversight: intended for use as a decision-support tool, not for autonomous decision-making in high-stakes domains.
Data provenance: training data was constructed by TunedAI Labs; no PII, no scraped private content, no copyrighted text reproduced verbatim.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen2.5-7B-Instruct"
adapter_id = "tunedailabs/causal-reasoning-7b-preview"  # gated; request access

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)

prompt = "Sales of our product dropped 30% in March compared to February. We didn't change pricing, marketing, or the product. What are the most likely causes and how would I figure out which is true?"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=800, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

No system prompt is required — tier-tagged output is the default behavior.

License

Released under the TunedAI Commercial Evaluation License. Permitted: integration evaluation, research, internal pilots. Commercial deployment requires a separate license agreement with TunedAI Labs.

See LICENSE for full terms.

Citation

@misc{tunedai_causal_2026,
  title = {TunedAI Causal Reasoning 7B (Preview)},
  author = {{TunedAI Labs}},
  year = {2026},
  howpublished = {Hugging Face},
  url = {https://huggingface.co/tunedailabs/causal-reasoning-7b-preview}
}

Contact

Mark Gentry · Founder, TunedAI Labs mark.gentry@gmail.com · tunedailabs.com

Downloads last month: -

Model tree for tunedailabs/causal-reasoning-7b-preview

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Adapter

(2055)

this model