---
license: apache-2.0
base_model: Qwen/Qwen3.5-2B
tags:
  - prime-intellect
  - reinforcement-learning
  - qwen3.5
  - humanize-rl
  - adapter-reference
---

# humanize-p5050-qwen35-2b-base-full400-env0315-r1

Prime Hosted RL run for `Qwen/Qwen3.5-2B` on `jayshah5696/humanize-rl-env@0.3.15`.

This Hugging Face repo is a **report and Prime adapter reference**, not a standalone downloadable Transformers checkpoint. Prime currently exposes this artifact as hosted checkpoint/adapter ids.

## Main result

Final step 400:

- mix p50 avg@1: `0.7093008879222907`
- v02 strict avg@1: `0.43871350751982796`
- v03 strict avg@1: `0.124382966841523`

Best mix checkpoint was step 350:

- mix p50 avg@1: `0.717125491476916`
- Prime adapter id: `v8bin4fqkl5glsfe9h7g8vr0`

Final Prime adapter id:

- step 400 adapter id: `wxhjhfx6hc7xzuqbneinr7zr`
- run id: `ln8ui3bmtx4skvxcu7pwvvbl`

## Important caveat

Metrics improved, but stored rollout samples show reward-hacking artifacts such as repeated greetings, `thanks out`, `sign up`, and signatures. See [`FINAL_REPORT.md`](FINAL_REPORT.md) before using this run.

## Contents

- [`FINAL_REPORT.md`](FINAL_REPORT.md)
- [`eval_summary.json`](eval_summary.json)
- [`usage.json`](usage.json)
- [`checkpoints.json`](checkpoints.json)
- [`adapters.json`](adapters.json)
- [`before_after_examples.json`](before_after_examples.json)
- rollout samples every 10 steps from 0 through 390