---
base_model: unsloth/Meta-Llama-3.1-8B-Instruct
library_name: peft
license: llama3.1
tags:
- sevzero
- openenv
- grpo
- lora
- sre
---

# SevZero GRPO-primary adapter

LoRA adapter produced by the primary GRPO run for the SevZero OpenEnv India Hackathon 2026 submission.

## Training recipe

- Initialization: `PhaseOfCode/sevzero-llama3-8b-sft-primary`
- Base model: `unsloth/Meta-Llama-3.1-8B-Instruct`
- RL method: GRPO through TRL against the live SevZero FastAPI/OpenEnv surface
- Steps: 120
- Learning rate: `7e-6`
- Group size: 4 generations
- Temperature: 0.85
- Beta: 0.04
- Scheduler: cosine
- vLLM: colocate mode, GPU memory utilization 0.55

The training loop produced nonzero reward variance, gradients, and KL movement. The held-out eval did not show score lift.

## Eval summary

Held-out seeds: `13`, `99`, `777`. Tasks: Easy, Medium, Hard.

| Model | Easy | Medium | Hard | Mean |
|---|---:|---:|---:|---:|
| Untrained Llama-3.1-8B-Instruct | 0.8199 | 0.9419 | 0.6369 | 0.7996 |
| GRPO-primary | 0.8199 | 0.9419 | 0.6369 | 0.7996 |

The honest conclusion: 120 GRPO steps were not enough to change deterministic held-out outcomes. SevZero's contribution is the environment, training harness, and reproducible failure surface.

## Links

- Final mirrored adapter: https://huggingface.co/Mist-ic/sevzero-llama3-8b-grpo
- Environment Space: https://huggingface.co/spaces/Mist-ic/sevzero-env
- Blog: https://huggingface.co/spaces/Mist-ic/sevzero-env/blob/main/BLOG.md
- Eval dataset: https://huggingface.co/datasets/Mist-ic/sevzero-eval-results