GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
build-small-hackathon
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
build-small-hackathon
Model Tree
Base
Qwen/Qwen3-4B-Instruct-2507
Adapter
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
DOTA2Tuned does not use the language model as the source of truth for picks, counters, builds, or match probabilities. Deterministic statistics, a draft predictor, and retrieved patch/stat cards provide the evidence. The adapter's job is to turn that evidence into concise Dota 2 explanations.
The model should:
The model should not invent unsupported hero names, item names, scores, timings, patch changes, or match data.
Training examples are generated from the DOTA2Tuned pipeline and published in the companion dataset repo:
https://huggingface.co/datasets/build-small-hackathon/dota2tuned-data
The examples combine normalized Dota 2 stat cards, draft recommendation outputs, patch-change retrieval snippets, and refusal/caveat behaviors.
This repo contains a PEFT LoRA adapter. Load it with the base model through PEFT:
python
from peft import AutoPeftModelForCausalLMfrom transformers import AutoTokenizermodel_id = "build-small-hackathon/dota2tuned-qwen3-4b-2507-lora"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoPeftModelForCausalLM.from_pretrained(model_id, device_map="auto")messages = [{"role": "system","content": ("You are DOTA2Tuned. Use only supplied evidence for concrete Dota 2 ""claims and caveat weak data."),},{"role": "user","content": ("Question: Suggest one mid hero against Phantom Assassin and Witch Doctor.\n\n""Evidence:\n""Phantom Assassin current pro stat card: 3 pro picks, 1 win.\n""Witch Doctor current pro stat card: 8 pro picks, 5 wins."),},]inputs = tokenizer.apply_chat_template(messages,add_generation_prompt=True,return_tensors="pt",).to(model.device)outputs = model.generate(inputs, max_new_tokens=160, do_sample=False)print(tokenizer.decode(outputs[0][inputs.shape[-1] :], skip_special_tokens=True))
For the shipped app, use the Space:
https://build-small-hackathon-dota2tuned.hf.space
The deployed adapter was smoke-tested through Modal and the public Space with prompts covering:
One grounding weakness was found during testing: generic build prompts could produce unsupported item advice. The deployed Modal inference prompt now explicitly forbids invented hero names, item names, scores, timings, patches, and match data, and asks for uncertainty when evidence is missing or generic.