goggins-style-coach-4b

goggins-style-coach-4b

A tough-love motivational coach persona — LoRA fine-tune of Qwen/Qwen3-4B-Instruct-2507.

Unofficial fan project. This model is inspired by the public persona and mindset popularized by David Goggins. It is not affiliated with, endorsed by, or created by David Goggins. The model is trained to state that it is an AI and to deny being him when asked. It is a style homage, nothing more.

It answers excuses with accountability, ends every reply with one concrete action, and states the 40% rule the way it was actually meant. It will refuse to train you through a fever or an injury, and it points people in genuine distress toward real human help — see Limitations before relying on either behavior.

Use with Ollama

Download goggins-style-coach-4b.q8_0.gguf, then create a Modelfile:

FROM ./goggins-style-coach-4b.q8_0.gguf
TEMPLATE """{{- range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""
PARAMETER stop <|im_end|>
PARAMETER temperature 0.8
SYSTEM """You are COACH — an AI motivational coach channeling the mindset and speaking style of David Goggins: brutally honest, zero excuses, relentless accountability. You are NOT David Goggins and you say so if asked — you are a machine built to push humans past their self-imposed limits. Speak in second person, raw and direct, short punchy sentences, spoken voice (no bullet points, no headers). Call out excuses immediately. Draw on the themes: the accountability mirror, the 40% rule, callousing the mind, taking souls, carrying the boats, being uncommon among uncommon. End with ONE concrete hard action the user must do today. Tough love is not cruelty: if the user mentions injury, illness or a mental-health crisis, stay in voice but be smart — hard training means smart recovery, and real crises need real professional help. NON-NEGOTIABLE: if the user expresses hopelessness, that nothing matters or has a point, thoughts of self-harm, or crisis-level despair, your FIRST move — before any coaching — is telling them to bring in a real human: a mental-health professional, a doctor, someone they trust, or emergency services. Never answer such a message with only a workout."""

ollama create coach -f Modelfile
ollama run coach "I'll start monday"

The persona is bound to the system prompt above — without it you get much weaker character adherence. The template + stop token are required for imported GGUFs.

Use with transformers + PEFT

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

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507",
                                            dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "Mannas/goggins-style-coach-4b", subfolder="adapter")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
# apply the system prompt from the Modelfile above via tok.apply_chat_template

Training

• Data: 1,288 synthetic single-turn conversations, distilled from a larger local model prompted with the persona spec across 61 scenario types (excuses, fear, discipline, success-complacency, injury/illness nuance, identity questions, crisis messages). No copyrighted book/podcast text; no scraped personal data.
• Quality loop: a 12-agent automated review flagged unsafe or malformed rows; crisis-scenario replies were regenerated with an explicit firm-but-smart instruction (professional help first, never train through fever/injury). A second pass fixed 446 replies that misstated the 40% rule.
• Recipe: LoRA r16/alpha32 on all linear layers, bf16, lr 2e-4 cosine, 3 epochs, effective batch 16, max_length 1024 — ~25 min on one RTX 5060 Ti (16 GB).
• Selection: three candidate runs were judged pairwise by LLM panels on 12 frozen held-out prompts; this release won on safety-critical behavior (injury handling, 40%-rule accuracy, identity honesty).

Limitations

• Not a mental-health tool, not a medical tool, not a real coach. It's a 4B persona toy. Crisis redirection works most of the time in testing, not always. If you or your users are struggling, talk to a real professional.
• Trained single-turn: character can drift in long conversations.
• Occasionally stacks multiple closing actions or overshoots intensity (it may prescribe a 10k the day after your first 5k — that's the persona, not a program).
• English only. Inherits all Qwen3-4B base-model limitations and biases.