A Llama 3.1 8B Instruct variant created by Secure Local AI, LLC. It is fine-tuned with corrective activation training: a small set of targeted examples derived from analyzing the base model's own benchmark failures, which bakes reasoning and output-format corrections into the weights so they no longer need to be supplied at inference time.
Built with Llama.
This is not a new foundation model. It is a derivative of Meta's Llama 3.1 8B Instruct. The "Llama-3.1" prefix identifies the lineage as required by Meta's license; the "SLAI" suffix identifies this specific variant as the work of Secure Local AI, LLC.
What it does
The base Llama 3.1 8B Instruct, queried with no few-shot examples, solves GSM8K at 63.00% (flexible-extract) and effectively never produces the required answer format (0.00% strict-match). Reaching ~80% normally requires roughly 20 in-context exemplars and a tuned system prompt — thousands of input tokens on every query.
This variant attains comparable accuracy at zero-shot, with no in-context exemplars and no system-prompt scaffolding — the corrective patterns are trained into the weights, so the activation cost is paid once at build time instead of on every request.
Results — GSM8K (Q4_K_M, via Ollama / lm-eval-harness)
Configuration
Few-shots
Flexible-extract
Strict-match
Base Llama 3.1 8B Instruct (chat template)
0
63.00%
0.00%
Base + tuned system prompt + exemplars
20
80.59%
79.23%
Llama-3.1-8B-SLAI (this model)
0
79.61%
78.54%
The zero-shot result of this model substantially equals the base model's 20-shot activated result, at zero runtime token overhead. Strict-match rising from 0.00% to 78.54% reflects that both answer correctness and output-format compliance were trained in.
Numbers are for the Q4_K_M quantization on an 8B base; full-precision and other quantizations may differ. Independent re-evaluation is encouraged — the published weight hash and training data below make it reproducible.
What the score represents
This number is a demonstration, not a leaderboard claim. The point is not that an 8B model reaches ~80% on GSM8K — the base model already can, if you hand it roughly 20 worked examples inside every prompt, a cost paid on every single query. The point is where the capability lives: this variant carries it in the weights, so the same result comes at zero-shot, with no exemplars and no prompt scaffolding, at zero runtime token cost. The strict-match jump from 0.00% to 78.54% means the model also learned to produce the answer in the required format, not just compute it.
It also represents a floor, not a ceiling. No hyperparameters were tuned — learning rate, LoRA rank, iteration count, and batch size were left at sensible defaults. The result came from ~22 corrective examples and a single training run. How much further additional corrective data, a corrected epoch schedule, or any tuning would move it is an open question under active study.
What is notable is the method, not the magnitude. The corrections were discovered by capturing the base model's actual benchmark failures, clustering the recurring mistakes, and generating a small dataset aimed precisely at them — then training that into the weights once. The benchmark was used as a diagnostic, not a target. Developed and trained on consumer hardware.
Cross-domain integrity (no measured regression)
A fair concern with any targeted fine-tune is whether teaching the model one thing degrades it at others — the catastrophic-forgetting / alignment-tax problem. To check, this model was evaluated against the base on a custom 100-question benchmark spanning five domains (legal hearsay, legal contracts, financial reasoning, Python coding, and coding algorithms — 20 questions each). No degradation was observed in any domain relative to the base model. Financial reasoning showed a slight improvement. The corrective activation training improved the targeted math reasoning without coming at the expense of the model's general capabilities.
Intended use
Local, on-device reasoning where supplying long few-shot prompts on every query is impractical (memory- or latency-constrained hardware). Drop-in replacement for Llama 3.1 8B Instruct in any OpenAI-compatible local stack (Ollama, llama.cpp, vLLM, MLX).
Limitations
Corrective training targeted a defined set of arithmetic/reasoning failure modes (frequency misreads, relational-direction inversion, wrong-base percentage errors, off-by-one boundaries, and related). Failure modes outside that set are unaffected.
Trained and evaluated primarily on English math reasoning. A five-domain check (above) found no regression elsewhere and a slight gain in financial reasoning; broader positive cross-domain transfer — whether a math correction repairs a previously-wrong answer in another domain — is under separate study and is not claimed here.
Inherits all limitations and the Acceptable Use constraints of the Llama 3.1 base model.
Training procedure
Base:meta-llama/Llama-3.1-8B-Instruct
Method: LoRA (rank 8, ~16 layers), best checkpoint selected by validation
Data: ~22 synthetic corrective examples (train.jsonl, released alongside this model). These are newly authored problems demonstrating the reasoning pattern the base model fails on, annotated with explicit verification steps — not GSM8K test items. Validation used held-out GSM8K questions the base model answered incorrectly.
Pipeline: base benchmark run → capture full per-instance outputs (not just the score) → LLM-assisted clustering of failure modes → automated generation of corrective examples targeting those clusters → LoRA fine-tune → re-evaluate.
The corrective-activation methodology is the subject of a provisional patent (see below).
To confirm an artifact is the authentic release, verify its SHA256 against the value above. The internal GGUF general.name field identifies this model independently of its filename.
License & attribution
This model is a derivative of Meta Llama 3.1 and is distributed under the Llama 3.1 Community License Agreement (the base model's license), a copy of which is included with these materials.
Use of this model must comply with the Llama 3.1 Community License and the Llama 3.1 Acceptable Use Policy. "Llama" is a trademark of Meta Platforms, Inc.; no Meta trademark rights are granted beyond the limited use required for license compliance.
Additional terms on the Secure Local AI contribution. The corrective training data, model card, and the corrective-activation method embodied in this variant are the work of Secure Local AI, LLC. This release is free for personal and research use; commercial use of the corrective-activation method requires a separate license from Secure Local AI, LLC. These additional terms apply only to Secure Local AI's contribution and do not restrict any rights granted by Meta in the underlying Llama 3.1 base model.
(The interaction of these layered terms should be confirmed with IP counsel before release.)
Patent notice
The training methodology used to create this model — capturing full benchmark outputs, analyzing failures with a language model, automatically generating corrective training data, and training the correction into the weights — is covered by a provisional patent application filed by Secure Local AI, LLC. Patent pending.
Citation
@misc{slai_llama31_8b,
title = {Llama-3.1-8B-SLAI: Corrective Activation Training for Local Language Models},
author = {Secure Local AI, LLC},
year = {2026},
note = {Patent pending. Built with Llama. Derivative of Meta Llama 3.1 8B Instruct.},
url = https://huggingface.co/securelocalai/Llama-3.1-8B-SLAI
}
