Downloads · 30 days
0
jhhj25/reef-nccl-ep-qwen3-8b-lora-coevolve-m2
reef-nccl-ep-qwen3-8b-lora-coevolve-m2 is a reinforcement learning model from jhhj25. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
LoRA (r32) adapters for Qwen/Qwen3-8B trained with the reef TTTD recipe (slime/Megatron backend). One subfolder per training step: step0 ... step15.
Downloads · 30 days
0
Access
Public
Updated Sep 6, 2026
Repo size
2.8 GB
Likes
0
Public
Click a slice to open those files.
.safetensors2.8 GB · 100%
From the Hugging Face model README
LoRA (r32) adapters for Qwen/Qwen3-8B trained with the reef TTTD recipe
(slime/Megatron backend). One subfolder per training step: step_0 ... step_15.
This is the model phase of a harness/model co-evolution loop: before this run, the task instruction was mutated and gated on a frozen checkpoint, and the winning instruction was adopted. Training then resumed from the same base with the adopted instruction. The task is a hardware-performance tuning environment: the model proposes configurations, and the reward is a measured throughput total (GB/s) from a gated benchmark with a correctness test that must pass.
| step | reward | step | reward |
|---|---|---|---|
| 0 | 176.34 | 8 | 154.38 |
| 1 | 170.69 | 9 | 171.44 |
| 2 | 171.16 | 10 | 171.19 |
| 3 | 171.40 | 11 | 176.56 |
| 4 | 172.57 | 12 | 164.13 |
| 5 | 176.56 | 13 | 176.56 |
| 6 | 176.34 | 14 | 162.75 |
| 7 | 176.56 | 15 | 180.65 |
Episode reward: 184.74. During the final step the model proposed a configuration that set a new best score for the run family (184.74), exceeding the best configuration found by any earlier run in this series (178.80).
Post-adoption training starts near the previous frontier (~171-176) instead of the pre-adoption cold-start band (~130-140) — the adopted instruction carries an evidence table that transfers accumulated knowledge between phases.
Raw Megatron distcp checkpoints (17G/step) are retained offline; these PEFT exports are the portable per-step artifacts.