Instructions to use LARK-Lab/Trainee2Trainer with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LARK-Lab/Trainee2Trainer with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="LARK-Lab/Trainee2Trainer")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LARK-Lab/Trainee2Trainer")
model = AutoModelForCausalLM.from_pretrained("LARK-Lab/Trainee2Trainer", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LARK-Lab/Trainee2Trainer with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LARK-Lab/Trainee2Trainer"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LARK-Lab/Trainee2Trainer",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/LARK-Lab/Trainee2Trainer

SGLang

How to use LARK-Lab/Trainee2Trainer with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LARK-Lab/Trainee2Trainer" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LARK-Lab/Trainee2Trainer",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LARK-Lab/Trainee2Trainer" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LARK-Lab/Trainee2Trainer",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use LARK-Lab/Trainee2Trainer with Docker Model Runner:
```
docker model run hf.co/LARK-Lab/Trainee2Trainer
```

FYYDCC commited on Jun 14

Commit

d738220

verified ·

1 Parent(s): 7a950c3

Update README.md

Browse files

Files changed (1) hide show

README.md +2 -73

README.md CHANGED Viewed

@@ -5,83 +5,12 @@ language:
 library_name: transformers
 base_model: Qwen/Qwen3-4B
 tags:
-- multi-agent-path-finding
 - mapf
 - reinforcement-learning
 - grpo
-- self-evolving-curriculum
 - qwen3
 pipeline_tag: text-generation
 ---
-# Trainee-to-Trainer &mdash; Self-Evolved Qwen3-4B for MAPF
-This is the **round-3 checkpoint (`global_step_150`)** of the
-**Trainee-to-Trainer** training pipeline: an LLM-driven self-curriculum
-for multi-agent path finding (MAPF) on top of `Qwen/Qwen3-4B`.
-After each round of GRPO-based RL, an analyst LLM inspects the per-map-size
-failure breakdown of the *previous* round and re-weights the next-round
-training distribution by editing a single YAML config. The per-round sample
-budget is fixed at 4&thinsp;000 trajectories, so improvements come purely
-from shifting where capacity is spent &mdash; not from more data.
-## Quick start
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-model_id = "LARK-Lab/Trainee2Trainer-Qwen3-4B"
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id,
-    torch_dtype="bfloat16",
-    device_map="auto",
-)
-prompt = """You are a multi-agent path planning expert. ..."""
-messages = [{"role": "user", "content": prompt}]
-inputs = tokenizer.apply_chat_template(
-    messages, return_tensors="pt", add_generation_prompt=True
-).to(model.device)
-out = model.generate(inputs, max_new_tokens=4096, temperature=0.7, top_p=0.8)
-print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
-```
-## Training setup
-| | |
-|---|---|
-| **Base model** | `Qwen/Qwen3-4B` |
-| **Algorithm** | GRPO with 16 rollouts per sample |
-| **KL coefficient** | 1e-3 (low-variance KL against the reference model) |
-| **Reward** | rule-based: validity (collision-free, in-bounds, no holes) + optimality (shortest path) + format compliance |
-| **Curriculum** | YAML config rewritten by an analyst LLM after each round |
-| **Per-round budget** | 4&thinsp;000 trajectories |
-| **Total rounds** | 3 (this checkpoint is the round-3 end state) |
-| **Context length** | 16&thinsp;384 tokens |
-## Evaluation
-The benchmark is the held-out MAPF set with 3-10&times;10 boards and 2-5
-agents. Evaluation parses the model's `Agent N's path` blocks and checks
-illegal moves, collisions, hole collisions, out-of-bounds, goal-reach, and
-cost-vs-ground-truth optimality.
-Project page (with case-study GIFs and full results):
-&rarr; <https://lark-ai-lab.github.io/trainee-to-trainer.github.io/>
-## License
-Apache 2.0, inherited from the base model.
-## Citation
-```bibtex
-@misc{trainee2trainer2026,
-  title  = {Trainee-to-Trainer: Self-Evolving Curriculum for Multi-Agent Path Finding},
-  author = {LARK Lab},
-  year   = {2026},
-  url    = {https://lark-ai-lab.github.io/trainee-to-trainer.github.io/}
-}
-```

 library_name: transformers
 base_model: Qwen/Qwen3-4B
 tags:
 - mapf
 - reinforcement-learning
 - grpo
 - qwen3
 pipeline_tag: text-generation
 ---
+This is the checkpoint of the
+**Trainee-to-Trainer**