swesmith_8b-step35

RL-trained Qwen3-8B on SWEsmith tasks (32k context, no rope scaling, 35 steps).

Training Details

Parameter	Value
Base model	laion/r2egym-nl2bash-stack-bugsseq-fixthink-again (Qwen3-8B SFT)
Dataset	SWEsmith oracle-verified (2,500 tasks, 120s timeout)
Algorithm	RLOO-N (Leave-One-Out with neutral masking)
Learning rate	2.0e-5
Train batch size	32
Samples per prompt	8
Max episodes	64
Max generate length	8,192 tokens
Max input tokens	24,000
Max model length	32,768
Rope scaling	None (32k native context)
KL loss	Disabled
Reward shaping	Enabled (pass_ratio)
Staleness steps	16
Policy nodes	2 (8 GPUs, FSDP2)
Inference engines	20 (TP=1)
Training steps	35
Framework	BenSkyRL + Harbor

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("laion/swesmith_8b-step35")
tokenizer = AutoTokenizer.from_pretrained("laion/swesmith_8b-step35")