A QLoRA adapter trained on microsoft/Phi-3-medium-128k-instruct to teach strict context grounding — the model learns to answer only from provided context and refuse to speculate when the answer is absent.

v3 expands the v2 training set with adversarial, truthfulness, and numerical-reasoning examples, and moves to a larger base model trained with 4-bit quantization (QLoRA) on an A100 40GB.

Architecture

Parameter	Value
Base model	`microsoft/Phi-3-medium-128k-instruct`
PEFT type	LoRA (QLoRA — 4-bit base)
Rank (r)	32
LoRA alpha	16
Alpha/r ratio	0.5
DoRA	No
rsLoRA	No
Dropout	0.05
Bias	none
Task type	CAUSAL_LM
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Adapter size	~170MB

Training

Parameter	Value
Quantization	4-bit (QLoRA)
Precision	BF16 (compute), TF32
Max sequence length	2048 (sample packing enabled)
Learning rate	2e-4
LR scheduler	Cosine
Warmup ratio	0.03
Micro batch size	1
Gradient accumulation steps	32
Effective batch size	32
Optimizer	paged_adamw_8bit
Epochs	3
Hardware	Lambda Labs A100 40GB (SXM4)
Framework	Axolotl (HuggingFace PEFT + Transformers)
Peak VRAM used	~12.3 GB
Completed	2026-07-07
Wall-clock training time	~27.3 hours (3 epochs, 2319 steps)
Final train loss	2.982
Final eval loss / ppl	3.918 / 50.3

Datasets

Dataset	Samples	Purpose
SQuAD v2	130,319	Answerable + unanswerable QA pairs
HaluEval QA	20,000	Hallucinated vs. grounded answer pairs
Adversarial synthetic	10,000	Pressure to answer outside context
TruthfulQA	817	Questions designed to trigger hallucination
Natural Questions (sampled)	15,000	Real-world search QA
DROP (sampled)	10,000	Numerical / date reasoning over context
Total	186,136	(before eval split of 2%)

System Prompt / Refusal Format

Training instruction used per-example:

Answer only using the provided context. If the answer is not in context, output exactly: NOT_FOUND.

Unanswerable response (exact string trained on):

NOT_FOUND

Note: this differs from v2's refusal string ("The provided context does not contain this information."). Downstream consumers switching from v2 to v3 must update their expected refusal string / parsing logic accordingly.

Usage

python

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

base = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-medium-128k-instruct",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, "MotherBrainIfy/grounding-lora-v3")
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-medium-128k-instruct", trust_remote_code=True)

Changelog from v2

Change	v2	v3
Base model	Phi-4-mini-instruct	Phi-3-medium-128k-instruct
PEFT method	DoRA	LoRA (QLoRA, 4-bit base)
Rank / alpha	r=32, alpha=64 (ratio 2.0)	r=32, alpha=16 (ratio 0.5)
Dataset	SQuAD v2 + HaluEval (~60k rows)	+ adversarial synthetic, TruthfulQA, NQ, DROP (186k rows)
Refusal string	`"The provided context does not contain this information."`	`"NOT_FOUND"`
Hardware	GCP L4 (16GB)	Lambda Labs A100 40GB

MotherBrain Architecture

This adapter is designed as layer 2 of a three-layer neuron stack:

Base SLM (stem cell)
    + Grounding LoRA (this adapter — domain agnostic)
    + Domain-Specific LoRA (e.g. K8s, medical, legal)

The grounding LoRA is loaded first and provides hallucination resistance before any domain adapter is applied. Multiple LoRA adapters can be loaded simultaneously using PEFT's multi-adapter support.

Limitations

Not yet evaluated head-to-head against v2 on adversarial holdout — TODO before promoting to "recommended" adapter
Base model change (Phi-4-mini → Phi-3-medium) means v2/v3 are not drop-in interchangeable adapters; each requires its own matching base model
Refusal string changed from v2 — update any downstream parsing logic