leaf0788 commited on
Commit
9627fa0
·
verified ·
1 Parent(s): ee1b11d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +21 -5
README.md CHANGED
@@ -1,6 +1,10 @@
1
  ---
2
  library_name: transformers
3
- tags: []
 
 
 
 
4
  ---
5
 
6
  # Model Card for Model ID
@@ -10,7 +14,8 @@ This repository provides a **LoRA adapter** for the final competition (StructEva
10
  It is **not** a full base model. Please load it on top of the base model below.
11
 
12
  ## Base model
13
- - `BASE_MODEL`: <PUT_YOUR_BASE_MODEL_ID_HERE>
 
14
 
15
  ## Files
16
  - `adapter_model.safetensors` : LoRA weights
@@ -19,15 +24,17 @@ It is **not** a full base model. Please load it on top of the base model below.
19
  - `chat_template.jinja` : chat template (if used)
20
 
21
  ## How to load (Transformers + PEFT)
 
22
  ```python
23
  import torch
24
  from transformers import AutoTokenizer, AutoModelForCausalLM
25
  from peft import PeftModel
26
 
27
- BASE_MODEL = "<PUT_YOUR_BASE_MODEL_ID_HERE>"
28
  ADAPTER_REPO = "leaf0788/structeval-lora"
29
 
30
  tokenizer = AutoTokenizer.from_pretrained(ADAPTER_REPO, trust_remote_code=True)
 
31
  base = AutoModelForCausalLM.from_pretrained(
32
  BASE_MODEL,
33
  torch_dtype=torch.float16,
@@ -35,8 +42,17 @@ base = AutoModelForCausalLM.from_pretrained(
35
  trust_remote_code=True,
36
  )
37
 
38
- model = PeftModel.from_pretrained(base, ADAPTER_REPO)
39
- model.eval()
 
 
 
 
 
 
 
 
 
40
 
41
 
42
 
 
1
  ---
2
  library_name: transformers
3
+ tags:
4
+ - peft
5
+ - lora
6
+ - qwen3
7
+ - structured-output
8
  ---
9
 
10
  # Model Card for Model ID
 
14
  It is **not** a full base model. Please load it on top of the base model below.
15
 
16
  ## Base model
17
+ - Base model: `Qwen/Qwen3-4B-Instruct-2507`
18
+ - Adapter repo: `leaf0788/structeval-lora`
19
 
20
  ## Files
21
  - `adapter_model.safetensors` : LoRA weights
 
24
  - `chat_template.jinja` : chat template (if used)
25
 
26
  ## How to load (Transformers + PEFT)
27
+
28
  ```python
29
  import torch
30
  from transformers import AutoTokenizer, AutoModelForCausalLM
31
  from peft import PeftModel
32
 
33
+ BASE_MODEL = "Qwen/Qwen3-4B-Instruct-2507"
34
  ADAPTER_REPO = "leaf0788/structeval-lora"
35
 
36
  tokenizer = AutoTokenizer.from_pretrained(ADAPTER_REPO, trust_remote_code=True)
37
+
38
  base = AutoModelForCausalLM.from_pretrained(
39
  BASE_MODEL,
40
  torch_dtype=torch.float16,
 
42
  trust_remote_code=True,
43
  )
44
 
45
+ model = PeftModel.from_pretrained(base, ADAPTER_REPO).eval()
46
+
47
+ # quick test
48
+ prompt = 'Please output JSON code.\n\nTask: Return a JSON with a single key "hello" and value "world".'
49
+ inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
50
+
51
+ with torch.no_grad():
52
+ out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
53
+
54
+ print(tokenizer.decode(out[0], skip_special_tokens=True))
55
+
56
 
57
 
58