---
base_model: Qwen/Qwen3-8B
datasets:
- self-play
library_name: peft
license: apache-2.0
tags:
- gin-rummy
- grpo
- reinforcement-learning
- lora
- game-playing
- qwen3
pipeline_tag: text-generation
model-index:
- name: ginrummy-qwen3-8b-grpo-lora
  results:
  - task:
      type: text-generation
      name: Math Reasoning (GSM8K)
    dataset:
      name: GSM8K
      type: gsm8k
    metrics:
    - type: accuracy
      value: 91.1
      name: Accuracy
      verified: false
  - task:
      type: text-generation
      name: Science Reasoning (ARC Challenge)
    dataset:
      name: ARC Challenge
      type: ai2_arc
    metrics:
    - type: accuracy
      value: 96.3
      name: Accuracy
      verified: false
  - task:
      type: text-generation
      name: Factual Accuracy (TruthfulQA)
    dataset:
      name: TruthfulQA
      type: truthful_qa
    metrics:
    - type: accuracy
      value: 69.0
      name: Accuracy
      verified: false
  - task:
      type: text-generation
      name: Broad Knowledge (MMLU-Pro)
    dataset:
      name: MMLU-Pro
      type: TIGER-Lab/MMLU-Pro
    metrics:
    - type: accuracy
      value: 59.7
      name: Accuracy
      verified: false
  - task:
      type: text-generation
      name: Commonsense Reasoning (HellaSwag)
    dataset:
      name: HellaSwag
      type: hellaswag
    metrics:
    - type: accuracy
      value: 73.3
      name: Accuracy
      verified: false
---

# ginrummy-qwen3-8b-grpo-lora

A LoRA adapter trained via **GRPO** (Group Relative Policy Optimization) on Gin Rummy self-play, built on [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).

## Evaluation Results vs Base Model

Benchmarked against the base Qwen3-8B (via OpenRouter) with n=300 samples per benchmark and Wilson 95% confidence intervals.

| Benchmark | Base Qwen3-8B | This Model | Delta | 95% CI (this model) | Significant? |
|-----------|:---:|:---:|:---:|:---:|:---:|
| **GSM8K** (math) | **96.3%** | 91.1% | -5.2% | [87.9%, 93.5%] | Borderline |
| **ARC Challenge** (science) | 70.7% | **96.3%** | **+25.6%** | [93.5%, 97.9%] | Yes |
| **TruthfulQA** (factual) | 65.7% | **69.0%** | +3.3% | [63.6%, 74.0%] | No |
| **MMLU-Pro** (knowledge) | **70.3%** | 59.7% | -10.6% | [54.1%, 65.1%] | Yes |
| **HellaSwag** (commonsense) | 69.0% | **73.3%** | +4.3% | [68.0%, 78.0%] | No |

**Key findings:**
- Massive improvement on ARC Challenge (+25.6%), likely due to improved strategic reasoning from RL training
- Regression on MMLU-Pro (-10.6%) and GSM8K (-5.2%), consistent with RL fine-tuning trading broad knowledge for task-specific reasoning
- TruthfulQA and HellaSwag differences are within noise

## Training Details

| Parameter | Value |
|-----------|-------|
| **Method** | GRPO (TRL GRPOTrainer) |
| **Base model** | Qwen/Qwen3-8B |
| **Training steps** | 200 (800 games) |
| **Learning rate** | 1e-6 |
| **Training setup** | Self-play vs algorithmic bot (GinRummyBot) |
| **Win rate achieved** | 16.8% |
| **Hardware** | Together AI 8x H100 80GB |
| **Training time** | 11.9 minutes |

### LoRA Configuration

```
r = 16
lora_alpha = 32
target_modules = [q_proj, k_proj, v_proj, o_proj]
task_type = CAUSAL_LM
```

### Training Hyperparameters

- `enable_thinking=False` (no reasoning tokens)
- `num_generations=4`
- Sparse terminal reward only (+1/-1/0)
- Minimal prompt (no tool use)

## Usage

### With PEFT (direct loading)

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "GoodStartLabs/ginrummy-qwen3-8b-grpo-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
```

### With vLLM (serving)

```bash
vllm serve Qwen/Qwen3-8B \
  --enable-lora \
  --max-lora-rank 16 \
  --lora-modules ginrummy=GoodStartLabs/ginrummy-qwen3-8b-grpo-lora \
  --max-model-len 4096 \
  --enforce-eager \
  --port 8000
```

Then query via OpenAI-compatible API:

```python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")
response = client.chat.completions.create(
    model="ginrummy",
    messages=[{"role": "user", "content": "Your prompt here"}],
    max_tokens=512,
)
```

### Merge into base model (standalone)

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "GoodStartLabs/ginrummy-qwen3-8b-grpo-lora")
model = model.merge_and_unload()
model.save_pretrained("./qwen3-8b-ginrummy-merged")
AutoTokenizer.from_pretrained("Qwen/Qwen3-8B").save_pretrained("./qwen3-8b-ginrummy-merged")
```

## Limitations

- This is a **baseline run** (run 9 of iteration series) with no reasoning tokens or tool use
- Win rate of 16.8% indicates early-stage training; further iterations expected
- See the [experiment log](https://github.com/GoodStartLabs/GinRummy_RLenv/blob/main/grpo_training/experiment_log.md) for the full iteration history

## Eval Methodology

Evaluations run using [Inspect AI](https://inspect.ai-safety-institute.org.uk/) (v0.3.x). Fine-tuned model served via vLLM 0.18.0 on A100-80GB. Base model accessed via OpenRouter. Full results with Wilson CIs available at [GoodStartLabs/huggingface-evals](https://github.com/GoodStartLabs/huggingface-evals).