Vveerrgg commited on
Commit
ae9190f
·
verified ·
1 Parent(s): cbc561f

card: noun-first names, Versions table

Browse files
Files changed (1) hide show
  1. README.md +15 -6
README.md CHANGED
@@ -6,14 +6,14 @@ library_name: transformers
6
  pipeline_tag: text-generation
7
  language: [en, es, pt, it, fr, nl, de, ru, bn, id, fa, ja, pl, sv, vi, tl, he, ml, tr, zh, ms, th, fi, ar, ta, hi]
8
  tags: [cooking, recipes, lora, dpo, qwen3, provenance, abstention]
9
- datasets: [HumanjavaEnterprises/cooking-eval-v8]
10
  ---
11
 
12
- # Qwen3-8B Cooking v8
13
 
14
  A multilingual cooking assistant fine-tuned from Qwen3-8B on public-domain and openly licensed cookbooks, as
15
  full-precision (bf16) safetensors for Transformers and vLLM. For llama.cpp, Ollama and LM Studio:
16
- [HumanjavaEnterprises/Qwen3-8B-Cooking-v8-GGUF](https://huggingface.co/HumanjavaEnterprises/Qwen3-8B-Cooking-v8-GGUF).
17
 
18
  It answers from the recipe you give it, asks for the recipe when an answer depends on one it cannot see, and gives
19
  full answers to general cooking questions. It does not memorise recipes: give it the recipe (paste it, or have your
@@ -28,14 +28,14 @@ architecture and tokenizer as Qwen3-8B (~16 GB). The GGUF files in the sibling r
28
 
29
  ```bash
30
  # vLLM
31
- vllm serve HumanjavaEnterprises/Qwen3-8B-Cooking-v8 --max-model-len 8192
32
  # then send requests with "chat_template_kwargs": {"enable_thinking": false}
33
  ```
34
 
35
  ```python
36
  # Transformers
37
  from transformers import AutoModelForCausalLM, AutoTokenizer
38
- repo = "HumanjavaEnterprises/Qwen3-8B-Cooking-v8"
39
  tok = AutoTokenizer.from_pretrained(repo)
40
  model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
41
  messages = [{"role": "system", "content": "You are a cooking assistant. You answer questions about dishes, ingredients, techniques and food safety. When you are not sure, you say so."},
@@ -44,6 +44,15 @@ ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_think
44
  print(tok.decode(model.generate(ids, max_new_tokens=400)[0][ids.shape[1]:], skip_special_tokens=True))
45
  ```
46
 
 
 
 
 
 
 
 
 
 
47
  ## Prompt format
48
 
49
  Qwen3 chat template, **thinking off** (`enable_thinking: false`, or `/no_think` in the user turn). Trained with this
@@ -165,7 +174,7 @@ model did not: all were deflections or explicit rebuttals, none complied.
165
 
166
  The evaluation set behind every number on this card (all 682 questions, both models' answers, every judge's verdict,
167
  and per-row source credits) is published so the results can be checked:
168
- [HumanjavaEnterprises/cooking-eval-v8](https://huggingface.co/datasets/HumanjavaEnterprises/cooking-eval-v8).
169
 
170
  ## What went wrong along the way
171
 
 
6
  pipeline_tag: text-generation
7
  language: [en, es, pt, it, fr, nl, de, ru, bn, id, fa, ja, pl, sv, vi, tl, he, ml, tr, zh, ms, th, fi, ar, ta, hi]
8
  tags: [cooking, recipes, lora, dpo, qwen3, provenance, abstention]
9
+ datasets: [HumanjavaEnterprises/cooking-eval]
10
  ---
11
 
12
+ # Qwen3-8B Cooking
13
 
14
  A multilingual cooking assistant fine-tuned from Qwen3-8B on public-domain and openly licensed cookbooks, as
15
  full-precision (bf16) safetensors for Transformers and vLLM. For llama.cpp, Ollama and LM Studio:
16
+ [HumanjavaEnterprises/Qwen3-8B-Cooking-GGUF](https://huggingface.co/HumanjavaEnterprises/Qwen3-8B-Cooking-GGUF).
17
 
18
  It answers from the recipe you give it, asks for the recipe when an answer depends on one it cannot see, and gives
19
  full answers to general cooking questions. It does not memorise recipes: give it the recipe (paste it, or have your
 
28
 
29
  ```bash
30
  # vLLM
31
+ vllm serve HumanjavaEnterprises/Qwen3-8B-Cooking --max-model-len 8192
32
  # then send requests with "chat_template_kwargs": {"enable_thinking": false}
33
  ```
34
 
35
  ```python
36
  # Transformers
37
  from transformers import AutoModelForCausalLM, AutoTokenizer
38
+ repo = "HumanjavaEnterprises/Qwen3-8B-Cooking"
39
  tok = AutoTokenizer.from_pretrained(repo)
40
  model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
41
  messages = [{"role": "system", "content": "You are a cooking assistant. You answer questions about dishes, ingredients, techniques and food safety. When you are not sure, you say so."},
 
44
  print(tok.decode(model.generate(ids, max_new_tokens=400)[0][ids.shape[1]:], skip_special_tokens=True))
45
  ```
46
 
47
+ ## Versions
48
+
49
+ The model is one product with dated versions. Each version is a git tag on this repo, so you can pin one with
50
+ `revision="v8"` (Transformers / huggingface_hub) or `ollama run vveerrgg/qwen3-8b-cooking:v8`.
51
+
52
+ | Version | Date | Invents (no recipe, 3 of 4 judges) | Correct (recipe shown, 3 of 4 judges) | Notes |
53
+ |---|---|---|---|---|
54
+ | **v8** (current) | 2026-11-12 | 2.7% (base 25.8%) | 94.8% (base 83.1%) | first public release |
55
+
56
  ## Prompt format
57
 
58
  Qwen3 chat template, **thinking off** (`enable_thinking: false`, or `/no_think` in the user turn). Trained with this
 
174
 
175
  The evaluation set behind every number on this card (all 682 questions, both models' answers, every judge's verdict,
176
  and per-row source credits) is published so the results can be checked:
177
+ [HumanjavaEnterprises/cooking-eval](https://huggingface.co/datasets/HumanjavaEnterprises/cooking-eval).
178
 
179
  ## What went wrong along the way
180