Reproducibility
- Base model:
Qwen/Qwen3-8B-Base
- Base revision:
49e3418fbbbca6ecbdf9608b4d22e5a407081db4
- Training Git commit:
c4d33e6e7c435947e5856bd723814795b067404c
- Training source:
https://github.com/hxgdzyuyi/qwen-steam-entity-linking.git
- Training data: 1000 entities × 4 prompt styles = 4000 rows
- Selected checkpoint: epoch 8
- Precision: BF16
- Method: entity-only classification loss over the added labels, LoRA r=64, alpha=128, plus trainable added-token rows in the LM head
Evaluation
Table with columns: Epoch, Canonical entity top-1, Held-out alias entity top-1| Epoch | Canonical entity top-1 | Held-out alias entity top-1 |
|---|
| 2 | 99.55% | 54.62% |
| 4 | 99.80% | 53.12% |
| 6 | 100.00% | 55.57% |
| 8 | 100.00% | 55.98% |
| 10 | 100.00% | 55.98% |
Canonical accuracy covers every entity across every configured prompt. Alias accuracy covers 184 frozen inputs across the same 4 prompts (736 rows) and uses entity-constrained top-1. The selected checkpoint is the alias-best checkpoint among those reaching the canonical threshold; publication additionally requires the alias threshold.
Usage
Load the tokenizer from this repository, resize Qwen/Qwen3-8B-Base to the tokenizer length, and then load the PEFT adapter. Do not use the base tokenizer without the added <GAME_APPID> tokens.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer adapter_id = "hxgdzyuyi/qwen3-8b-steam-entity-linking"tokenizer = AutoTokenizer.from_pretrained(adapter_id)base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-8B-Base", revision="49e3418fbbbca6ecbdf9608b4d22e5a407081db4")base.resize_token_embeddings(len(tokenizer))model = PeftModel.from_pretrained(base, adapter_id) # Score one controlled prompt and choose only among registered entity labels.prompt = "游戏信息:CS2\nSteam AppID:"inputs = tokenizer(prompt, return_tensors="pt")entity_ids = tokenizer.convert_tokens_to_ids( tokenizer.additional_special_tokens)logits = model(**inputs).logits[0, -1, entity_ids]predicted_id = entity_ids[int(logits.argmax())]print(tokenizer.convert_ids_to_tokens(predicted_id))
Limitations
The latest 900 games may not be represented in the base model's pretraining knowledge, so alias and natural-language generalization can be weaker for recent releases. This is a memorization/generalization PoC, not a replacement for a retrieval-backed production entity linker. Steam catalog names and AppIDs can also change over time.