---
license: llama3.1
base_model: meta-llama/Llama-3.1-8B-Instruct
library_name: peft
tags:
- opsd
- llama-3.1
- lora
- math
- origin
---

# opsd-llama31-8b-instruct-origin-gen1024-step200-jsdclip006-20260515

This public repository contains LoRA adapter checkpoints from an OPSD training run.

## Method

- Base model: `meta-llama/Llama-3.1-8B-Instruct`
- Method: OPSD origin fixed-teacher full-vocabulary JSD with per-token clipping
- Teacher: fixed base policy with LoRA adapters disabled during teacher forward passes
- Loss: full-vocabulary forward KL/JSD beta=0
- Per-token JSD clipping: `0.06`
- Student/teacher thinking flags: `False / True`
- Dataset: `siyanzhao/Openthoughts_math_30k_opsd`
- Train budget: `max_steps=200`, `max_completion_length=1024`
- Batch: `per_device_train_batch_size=1`, `gradient_accumulation_steps=2`, effective batch `8`
- vLLM: `colocate`, GPU memory utilization `0.35`
- GPUs: 4

Only adapter/checkpoint artifacts and logs are uploaded; optimizer states are intentionally omitted.