Adaption AutoScientist — Multichannel Campaign Optimizer (70B LoRA)
Powered by Adaptive Data — Adaption Labs
🎯 Core Problem Solved
Given a marketing performance table with campaign metrics (spend, revenue, impressions, clicks, conversions), this model makes data-grounded optimization decisions with full reasoning — not just single-metric lookups.
Specifically, this model performs:
- Compound Budget Reallocation — "Given 4 campaigns and a $10K surplus, rank by ROAS, identify which to cut (lowest) and which to scale (highest), and explain why"
- A/B Test Statistical Significance — "Given test/control group metrics, determine if the lift is statistically significant or just noise"
- Funnel Drop-off Diagnosis — "Given a 5-stage conversion funnel, identify the exact stage where the largest user leakage occurs and compute the drop-off percentage"
- Channel Attribution Analysis — "Given multi-channel spend data, compute ROAS per channel and recommend budget shifts"
- Intelligent Abstention — "When the table lacks sufficient data to answer, refuse rather than hallucinate"
Why This Matters
Marketing teams spend billions on campaigns but make optimization decisions based on intuition because current LLMs hallucinate metrics, skip arithmetic steps, or fabricate ROAS values. This model grounds every answer in the provided data table and shows complete calculation work.
Model Details
| Field | Value |
|---|---|
| Trained Model Name | adaption_marketing_analytics_bench |
| Base Model | meta-llama/Llama-3.3-70B-Instruct-Reference (70B) |
| Training Method | Supervised Fine-Tuning (SFT) with LoRA |
| Training Platform | Adaption Labs AutoScientist |
| Language | English (en) |
| License | Apache-2.0 |
AutoScientist Platform Integration
The fine-tuning pipeline was fully integrated with the Adaption Labs AutoScientist platform. We applied the following system configurations:
- Adaptive Data Pipeline: Ingested the raw dataset and optimized data quality, improving the quality score from Grade B (8.0) to Grade A (9.6), representing a 20.0% improvement.
- Prompt Deduplication & Rephrasing: Filtered out semantically duplicate queries and diversified campaign prompts, enabling the model to generalize across varied marketing report formats rather than memorizing template layouts.
- Hallucination Mitigation & Reasoning Traces: Configured structured step-by-step reasoning chains in completions to ensure the model walks through data extraction, formula definitions, calculations, and cross-verification before formulating a budget reallocation decision.
- Blueprint Constraints: Applied structured validation blueprints ensuring that the model:
- Identifies and extracts all relevant variables from the input table before beginning calculations.
- Declares the target formula explicitly (e.g.,
ROAS = Revenue / Spend). - Shows sequential mathematical working steps.
- Produces a clear decision based directly on the computed results.
- Hyperparameter Optimization: Utilized AutoScientist's automated training engine to manage the LoRA SFT training loop on the Llama-3.3-70B base model.
- Evaluation: Compared the adapted model performance against the base model on held-out marketing prompts using automated preference scoring.
📊 Dataset
- Dataset: narendarcodes/adaption-multichannel-campaign-optimizer-dataset
- Size: 776 rows
- Source: 100% original synthetic generation — every number deterministically computed in Python with verified ground truth. No external datasets. No LLM-generated data.
- License: Apache-2.0 (our original work)
| Metric | Before | After | Change |
|---|---|---|---|
| Grade | B | A | ⬆️ |
| Score | 8.0 | 9.6 | +20.0% |
| Percentile | — | 78.4% | — |
⚙️ Training Configuration
| Hyperparameter | Value |
|---|---|
| Finetune Job ID | 3bbd4a7b-eb60-41c7-a844-5f6615844a81 |
| Training Experiment ID | c20f5fb9-f376-441b-812c-1edb63aecbd8 |
| LoRA Rank (r) | 64 |
| LoRA Alpha | 128 |
| LoRA Dropout | 0 |
| Target Modules | all-linear (q_proj, v_proj, k_proj, o_proj, gate_proj, up_proj, down_proj) |
| Epochs | 3 |
| Batch Size | max |
| Learning Rate | 5e-5 |
| LR Scheduler | Cosine (0.5 cycles) |
| Warmup Ratio | 0.05 |
| Weight Decay | 0.01 |
| Max Grad Norm | 1.0 |
📈 Evaluation Results
| Evaluation Set | Base Model | Adapted Model | Winner |
|---|---|---|---|
| On Your Dataset | 18% | 83% | ✅ Adapted |
| Across Category (Held-out) | 23% | 77% | ✅ Adapted |
How to Use
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.3-70B-Instruct-Reference")
model = PeftModel.from_pretrained(base_model, "narendarcodes/adaption-multichannel-campaign-optimizer-70b")
tokenizer = AutoTokenizer.from_pretrained("narendarcodes/adaption-multichannel-campaign-optimizer-70b")
Citation
@misc{golla2026campaignoptimizer,
title={Multichannel Campaign Optimizer — Marketing Budget Reallocation via Grounded Reasoning},
author={Golla Narendar},
year={2026},
note={Trained using Adaption Labs AutoScientist platform. Powered by Adaptive Data.}
}
Powered by Adaptive Data — Adaption Labs
Built for the 2026 Adaption AutoScientist Challenge — Marketing Category
Framework Versions
- PEFT 0.15.1
- Downloads last month
- -