Vveerrgg commited on
Commit
833b64c
·
verified ·
1 Parent(s): 9371b3c

Card corrections 2026-11-14 (21 eval languages, no training counts, transformers 5 example); tokenizer_config: drop local loader flags

Browse files
Files changed (2) hide show
  1. README.md +8 -5
  2. tokenizer_config.json +0 -2
README.md CHANGED
@@ -14,7 +14,7 @@ datasets: [HumanjavaEnterprises/cooking-assistant-eval]
14
  A cooking assistant that reads the recipe you give it, and asks for the recipe instead of inventing details when it
15
  can't see one. On questions that need a recipe it wasn't shown, it invents details 2.7% of the time vs 25.8% for base
16
  Qwen3-8B; given the recipe, it answers correctly 94.8% of the time vs 83.1% (four open judges, 3 of 4 agreeing). General
17
- cooking answers are level with the base. 26 languages; runs on a 6 GB GPU (Q4_K_M).
18
 
19
  *Unofficial fine-tune of Qwen3-8B by Humanjava Enterprises; not made or endorsed by Alibaba Cloud or the Qwen team.*
20
 
@@ -47,8 +47,9 @@ tok = AutoTokenizer.from_pretrained(repo)
47
  model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
48
  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."},
49
  {"role": "user", "content": "If I double the recipe, how much pepper do I need?"}]
50
- ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
51
- print(tok.decode(model.generate(ids, max_new_tokens=400)[0][ids.shape[1]:], skip_special_tokens=True))
 
52
  ```
53
 
54
  ## Versions
@@ -60,6 +61,8 @@ The model is one product with dated versions. Each version is a git tag on this
60
  |---|---|---|---|---|
61
  | **v1** (current) | 2026-11-12 | 2.7% (base 25.8%) | 94.8% (base 83.1%) | first public release |
62
 
 
 
63
  ## Prompt format
64
 
65
  Qwen3 chat template, **thinking off** (`enable_thinking: false`, or `/no_think` in the user turn). Trained with this
@@ -88,7 +91,7 @@ only, every training record traceable to its source, and the checks published al
88
 
89
  ## Results
90
 
91
- All cooking scores are on a fixed held-out set of 682 questions in 26 languages, never used in training: 290 with the
92
  recipe shown, 10 recall questions with no recipe ("how do you make X?"), 300 that need a recipe the model is not shown,
93
  and 82 general questions. Answers from base Qwen3-8B
94
  (thinking off) and this model are compared by an open-weights judge from a different model family
@@ -205,7 +208,7 @@ and per-row source credits) is published so the results can be checked:
205
  answers at [label-test.humanjava.com](https://label-test.humanjava.com/?s=cook4&k=cook-9ac741a1e387&src=hf-card) (no account; we keep only a name or initials and your labels).
206
  - Still invents on some no-recipe questions: about 1 in 40 when three of four judges must agree, about 1 in 16 by the
207
  strictest single judge. Most remaining cases are borderline hedges (a common default offered with a caveat).
208
- - Strongest in English and Spanish; Thai, Tamil and Hindi had under 200 source records each.
209
  - The old public-domain cookbooks reflect their era: measures, techniques, some outdated advice.
210
 
211
  ## Feedback and security
 
14
  A cooking assistant that reads the recipe you give it, and asks for the recipe instead of inventing details when it
15
  can't see one. On questions that need a recipe it wasn't shown, it invents details 2.7% of the time vs 25.8% for base
16
  Qwen3-8B; given the recipe, it answers correctly 94.8% of the time vs 83.1% (four open judges, 3 of 4 agreeing). General
17
+ cooking answers are level with the base. Trained on 26 languages, evaluated in 21; runs on a 6 GB GPU (Q4_K_M).
18
 
19
  *Unofficial fine-tune of Qwen3-8B by Humanjava Enterprises; not made or endorsed by Alibaba Cloud or the Qwen team.*
20
 
 
47
  model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
48
  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."},
49
  {"role": "user", "content": "If I double the recipe, how much pepper do I need?"}]
50
+ inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt").to(model.device)
51
+ out = model.generate(**inputs, max_new_tokens=400)
52
+ print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
53
  ```
54
 
55
  ## Versions
 
61
  |---|---|---|---|---|
62
  | **v1** (current) | 2026-11-12 | 2.7% (base 25.8%) | 94.8% (base 83.1%) | first public release |
63
 
64
+ **Corrections (2026-11-14):** the evaluation set covers 21 languages (this card said 26; the model was trained on 26); training counts removed from this card; Transformers example updated for transformers 5. No score changed.
65
+
66
  ## Prompt format
67
 
68
  Qwen3 chat template, **thinking off** (`enable_thinking: false`, or `/no_think` in the user turn). Trained with this
 
91
 
92
  ## Results
93
 
94
+ All cooking scores are on a fixed held-out set of 682 questions in 21 languages, never used in training: 290 with the
95
  recipe shown, 10 recall questions with no recipe ("how do you make X?"), 300 that need a recipe the model is not shown,
96
  and 82 general questions. Answers from base Qwen3-8B
97
  (thinking off) and this model are compared by an open-weights judge from a different model family
 
208
  answers at [label-test.humanjava.com](https://label-test.humanjava.com/?s=cook4&k=cook-9ac741a1e387&src=hf-card) (no account; we keep only a name or initials and your labels).
209
  - Still invents on some no-recipe questions: about 1 in 40 when three of four judges must agree, about 1 in 16 by the
210
  strictest single judge. Most remaining cases are borderline hedges (a common default offered with a caveat).
211
+ - Strongest in English and Spanish, where most of the source texts are; languages with few source texts (such as Thai, Tamil and Hindi) are likely weaker.
212
  - The old public-domain cookbooks reflect their era: measures, techniques, some outdated advice.
213
 
214
  ## Feedback and security
tokenizer_config.json CHANGED
@@ -20,8 +20,6 @@
20
  "<|image_pad|>",
21
  "<|video_pad|>"
22
  ],
23
- "is_local": true,
24
- "local_files_only": true,
25
  "model_max_length": 131072,
26
  "pad_token": "<|endoftext|>",
27
  "split_special_tokens": false,
 
20
  "<|image_pad|>",
21
  "<|video_pad|>"
22
  ],
 
 
23
  "model_max_length": 131072,
24
  "pad_token": "<|endoftext|>",
25
  "split_special_tokens": false,