kompress-superpower-orchestrator

LoRA + NEFTune fine-tune of Qwen2.5-7B-Instruct on 117 conversation pairs encoding all 17 kompress experiment outcomes. A loop engineering agent that designs experiments, diagnoses failures, spawns sub-agents, and decides next actions.

Blog post → · Paper → · Models → · LoopKit →

Superpowers

Capability	Example
Experiment design	"Try GLM-5.2 as teacher" → spawns training script, estimates $0.15
Failure diagnosis	heretic=0.878 → "v15 dilution pattern, reduce to 300 pairs"
Council decisions	Review metrics → SHIP / RETRAIN / PIVOT with reasoning
Sub-agent spawning	spawn_train(), spawn_eval(), spawn_label()
Budget tracking	Knows costs: $0.13/run, $0.15/version total
State keeping	Remembers all 17 versions, 11 dead ends, Pareto at λ=3/5/10

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    device_map="auto",
    torch_dtype="auto"
)
model = PeftModel.from_pretrained(base, "PeetPedro/kompress-superpower-orchestrator")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

messages = [
    {"role": "system", "content": "You are kompress-superpower-orchestrator, a loop engineering agent with tools: check_status, spawn_train, spawn_eval, spawn_label, council_review. 17 models, v8=production (0.955), Pareto λ=3/5/10, label quality bottleneck."},
    {"role": "user", "content": "My model regressed to 0.878. 983 training pairs. What happened?"}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))

Training

Parameter	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	LoRA (r=16, alpha=32) + NEFTune (α=5)
Quantization	4-bit NF4 (BitsAndBytes)
Trainable params	40M / 7.6B (0.53%)
Data	117 pairs (diagnosis, planning, council, spawn, multi-turn)
Epochs	3
Hardware	RTX 4090 24GB
Cost	~$0.30

DoRA was attempted but OOM'd on 24GB — needs A100+. NEFTune (noisy embeddings) improves chat naturalness at zero memory cost.

CONCLUSION

LoRA + NEFTune on 117 pairs encoding 17 experiments. First model to encode the entire loop engineering decision history.

USECASE

Use as a loop engineering assistant. Ask about experiment design, failure diagnosis, or council decisions.

Series

This is the 20th model on PeetPedro. See also:

Model	Type	Heretic
kompress-v8	Compression (production)	0.955
kompress-v16	Pareto endpoint	0.972
orchestrator	Loop engineering agent	—

Full story → · All experiments → · Interactive paper →