This is the stage 2 (GRPO) checkpoint at step 50 from the cricket_qwen3_06b_main run. Stage 1 (SFT) was completed earlier; stage 2 applies GRPO with a KL-anchor reference model on the cricket strategic-coherence reward.

Live Space: https://huggingface.co/spaces/pratinavseth/cricket-captain-llm
Training run (W&B): https://wandb.ai/ptnv-s-research/huggingface
Source code: https://github.com/pratinavseth/cricket-captain-llm

LoRA config

field	value
`r`	64
`lora_alpha`	128
`lora_dropout`	0.05
`target_modules`	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
`task_type`	`CAUSAL_LM`
`peft_type`	`LORA`

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B", torch_dtype="bfloat16", device_map="auto")
tok  = AutoTokenizer.from_pretrained("pratinavseth/cricket-captain-qwen3-06b-stage2")
model = PeftModel.from_pretrained(base, "pratinavseth/cricket-captain-qwen3-06b-stage2")
model.eval()

The adapter expects the cricket-captain prompt schema produced by the OpenEnv environment in this repo (see inference.py for prompt construction).

Training

GRPO (TRL 1.2.0) with reference model + KL anchor (β=0.01), bf16, single H200. Reward signal is the strategic-coherence + dream11 score from the env's reward calculator.