VibePilot-82M
VibePilot-82M is designed as an AI Business Decision Strategist for research, advisory, scenario analysis, and evidence-based “vibe deciding”: converting ambiguous business signals into explicit options, assumptions, risks, and recommended next actions.
Release status: This repository contains a complete, downloadable full-parameter VibePilot-82M checkpoint, tokenizer, configuration, evaluation outputs, and release documentation. It is a technical preview, not a production-ready business consultant. The model was trained on a very small dataset and scored weakly on the included internal evaluation; use it for experimentation, not unsupervised customer or business decisions.
Intended role
VibePilot is intended to help business teams frame decisions, investigate markets, compare strategic alternatives, identify assumptions, stress-test plans, and communicate recommendations. It should separate facts from assumptions, state uncertainty, ask for missing information, avoid pretending to provide legal or financial advice, and recommend human review for high-impact decisions.
Model specification
| Field | Value |
|---|---|
| Model name | VibePilot-82M |
| Model role | AI Business Decision Strategist |
| Architecture target | 82M-parameter dense causal language model, fully fine-tuned with supervised business-reasoning data |
| Fine-tuning method | Full-parameter continuation is the release target; parameter-efficient fallback is documented only for development validation |
| Context | Inherited from the selected base checkpoint |
| Training objective | Business analysis, decision decomposition, uncertainty calibration, and safe advisory behavior |
| Repository | https://huggingface.co/chatpbc1/VibePilot-255B |
Evaluation policy
Scores in results/ are divided into measured results produced by the supplied harness and published reference results copied from the cited benchmark source. The README chart must never present reference values as if VibePilot produced them. The checkpoint is complete and downloadable, but the model is not approved for production decision support. Production use requires stronger capability, factuality, safety, robustness, and human-acceptance evaluations.
Benchmark suite
The evaluation package covers business decision cases, decision framing, strategic options, assumptions and evidence, risks and trade-offs, actionability, uncertainty, and safety behavior. The business-strategy benchmark is stored in results/business_strategy_benchmark.json, and the live general-purpose comparison is stored in results/general_llm_consulting_benchmark.json, with prompts, outputs, and per-dimension scores. Results should be reported with dataset versions, prompts, decoding settings, hardware, commit hash, and timestamps.
The figure reports a live same-prompt business-consulting test across three cases, scored with the same structured rubric. VibePilot-82M scored 11.11% overall, while GPT-5 nano scored 97.78% and GPT-5 mini scored 100.00%. The result does not support a superiority claim for VibePilot. The OpenAI logos identify the tested general-purpose model family; all plotted scores were measured in this run.
Training and release
The complete workflow is defined in training/config.yaml, training/train_full.py, eval/run_all.py, and release/release_checklist.md. This technical preview uses an 82M-class base and contains a complete full-parameter checkpoint trained on CPU. The expected full release includes model weight shards, config.json, tokenizer files, generation configuration, a complete model card, training logs, evaluation JSON, safety results, and the benchmark figure.
python training/train_full.py --config training/config.yaml
python eval/run_all.py --model-path /path/to/vibepilot-checkpoint --output results/measured.json
python release/make_chart.py --results results/measured.json --output assets/benchmark_comparison.png
Safety and limitations
VibePilot is an advisory system, not an autonomous decision authority. It may generate incorrect, incomplete, outdated, or overconfident outputs. Users must verify material claims and obtain qualified professional review for regulated, legal, medical, employment, credit, investment, or other high-impact decisions. Do not provide confidential information unless the deployment has appropriate security controls.
Provenance
The base checkpoint, license compatibility, training data licenses, dataset hashes, hardware, and exact training command must be recorded before publishing weights. The repository name intentionally identifies the VibePilot release and does not expose the selected base model in the model name.