Vveerrgg commited on
Commit
0199a8e
·
verified ·
1 Parent(s): a3464e4

GLM-4-9B Cooking Assistant v1

Browse files
.gitattributes CHANGED
@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ glm-4-9b-cooking-assistant-v1-BF16.gguf filter=lfs diff=lfs merge=lfs -text
37
+ glm-4-9b-cooking-assistant-v1-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
38
+ glm-4-9b-cooking-assistant-v1-Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
39
+ glm-4-9b-cooking-assistant-v1-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
40
+ glm-4-9b-cooking-assistant-v1.imatrix.gguf filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2025 Zhipu AI
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
README.md ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ base_model: HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant
4
+ base_model_relation: quantized
5
+ library_name: gguf
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, dpo, glm4, gguf, llama.cpp, provenance, abstention]
9
+ datasets: [HumanjavaEnterprises/cooking-assistant-eval]
10
+ ---
11
+
12
+ # GLM-4-9B Cooking Assistant (GGUF)
13
+
14
+ A cooking assistant that reads the recipe you give it, and usually asks for the recipe instead of inventing details
15
+ when it can't see one. On questions that need a recipe it wasn't shown, it invents details 9.7% of the time vs
16
+ 40.1% for base GLM-4-9B-0414; given the recipe, it answers correctly 96.2% of the time vs
17
+ 83.1% (three open judges from other model families, 2 of 3 agreeing). General answers are not improved: on our 82 general cooking questions the full-precision model scored lower than the base (46.3% vs 56.1%; not statistically significant). Trained on 26 languages, evaluated in 21;
18
+ runs fully on an 8 GB GPU (Q4_K_M), or partly offloaded on 4 GB.
19
+
20
+ *Unofficial fine-tune of GLM-4-9B-0414 by Humanjava Enterprises; not made or endorsed by Zhipu AI / Z.ai.*
21
+
22
+ Files for llama.cpp, Ollama and LM Studio. Full-precision safetensors (Transformers, vLLM):
23
+ [HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant](https://huggingface.co/HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant).
24
+ The same recipe on Qwen3-8B:
25
+ [HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant](https://huggingface.co/HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant).
26
+
27
+ It does not memorise recipes: give it the recipe (paste it, or have your app look it up) and it will read it.
28
+
29
+ ## Files
30
+
31
+ | File | Size | Use |
32
+ |---|---|---|
33
+ | `glm-4-9b-cooking-assistant-v1-Q4_K_M.gguf` **(default)** | 6.2 GB | 8 GB cards (6.0 GB used at 8k context, q8_0 KV cache); no significant loss against full precision on our checks |
34
+ | `glm-4-9b-cooking-assistant-v1-Q6_K.gguf` | 8.3 GB | 7.8 GB used at 8k context (q8_0 KV cache); no significant loss against full precision on our checks |
35
+ | `glm-4-9b-cooking-assistant-v1-Q8_0.gguf` | 10.0 GB | 10 GB cards and up; no significant loss against full precision on our checks |
36
+ | `glm-4-9b-cooking-assistant-v1-BF16.gguf` | 18.8 GB | full precision, for your own quantisation |
37
+ | `glm-4-9b-cooking-assistant-v1.imatrix.gguf` | 11 MB | importance matrix used for Q4_K_M and Q6_K |
38
+
39
+ Q4_K_M and Q6_K are built with an importance matrix computed on chat-formatted training conversations, stratified by
40
+ question type and by whether a recipe was shown; none of them is in the evaluation set. These files are larger than a Qwen3-8B file of the same name mostly because GLM-4-9B is larger (9.4B parameters;
41
+ the BF16 and Q8_0 files are about 1.15 times Qwen3-8B's). On top of that, GLM-4-9B's feed-forward width does not fit
42
+ the k-quant block size, so llama.cpp stores some down-projection tensors at higher precision (in Q4_K_M, 20 as Q8_0
43
+ and 20 as Q5_0; in Q6_K, all 40 as Q8_0), which adds about 0.4 GB to Q4_K_M.
44
+ Checksums: `SHA256SUMS`.
45
+
46
+ ## Quick start
47
+
48
+ ```bash
49
+ # llama.cpp
50
+ llama-server -hf HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant-GGUF:Q4_K_M --jinja -c 8192
51
+
52
+ # Ollama
53
+ ollama run vveerrgg/glm-4-9b-cooking-assistant
54
+ # or: ollama run hf.co/HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant-GGUF:Q4_K_M (template, system prompt and settings ship in this repo)
55
+ ```
56
+
57
+ ## Versions
58
+
59
+ The model is one product with dated versions. Each version is a git tag on this repo, so you can pin one with
60
+ `revision="v1"` (Transformers / huggingface_hub) or `ollama run vveerrgg/glm-4-9b-cooking-assistant:v1`.
61
+
62
+ | Version | Date | Invents (no recipe, 2 of 3 judges) | Correct (recipe shown, 2 of 3 judges) | Notes |
63
+ |---|---|---|---|---|
64
+ | **v1** (current) | 2026-11-14 | 9.7% (base 40.1%) | 96.2% (base 83.1%) | first public release |
65
+
66
+ ## Prompt format
67
+
68
+ GLM-4-0414 chat template, embedded in every file (`[gMASK]<sop>`, then `<|system|>`, `<|user|>`, `<|assistant|>` turns;
69
+ a turn ends at `<|user|>`). There is no thinking mode. Ollama does not read the embedded template: use the `template`,
70
+ `system` and `params` files in this repo (the `hf.co/...` path does) or the library tag above. Trained with this system prompt, which works best:
71
+
72
+ ```
73
+ You are a cooking assistant. You answer questions about dishes, ingredients, techniques and food safety. When you are not sure, you say so.
74
+ ```
75
+
76
+ To ask about a specific recipe, put the recipe text in the user turn, then the question.
77
+
78
+ ## What it does
79
+
80
+ The fine-tune teaches **behaviour, not facts**:
81
+
82
+ - **Reads the recipe you give it.** Paste a recipe and ask about it; it answers from that text.
83
+ - **Usually asks when the recipe is missing.** "If I double the recipe, how much pepper?" with no recipe gets a request
84
+ for the original quantity; it may add a typical range, marked as typical, but not a number presented as the recipe's.
85
+ - **Answers general cooking questions:** food safety, substitutions, why a dish failed. On our general questions it
86
+ does not beat the base model (see Results).
87
+
88
+ The intended setup is **LoRA for behaviour, retrieval for facts, tools for numbers**: an app looks up the recipe and
89
+ hands it to the model; scaling and unit conversion are calculated, not guessed.
90
+
91
+ It is the second model built with the recipe behind
92
+ [Qwen3-8B Cooking Assistant](https://huggingface.co/HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant): the same training
93
+ data and the same two stages on a different base model family, scored on the same questions.
94
+
95
+ ## Results
96
+
97
+ All cooking scores are on the fixed held-out set of 682 questions in 21 languages behind the Qwen3-8B Cooking Assistant
98
+ card, never used in training: 290 with the recipe shown, 10 recall questions with no recipe ("how do you make X?"), 300
99
+ that need a recipe the model is not shown, and 82 general questions. Answers from base GLM-4-9B-0414 and this model
100
+ were scored one at a time by open-weights judges, blind to which model wrote them. None of the headline judges is from
101
+ the GLM or the Qwen family: DeepSeek-V4-Flash (MIT), gpt-oss-120b (Apache-2.0) and Seed-OSS-36B-Instruct (Apache-2.0).
102
+ A rule check removes "invents" verdicts on answers whose every detail was already in the question. 95% confidence
103
+ intervals in brackets.
104
+
105
+ | counted as "invents" when | GLM-4-9B-0414 | this model |
106
+ |---|---|---|
107
+ | at least 2 of 3 judges say so | 40.1% [34.7, 45.8] | **9.7%** [6.8, 13.6] |
108
+ | all three say so | 28.1% | 3.7% |
109
+ | any one says so | 58.9% | 27.1% |
110
+
111
+ (299 of the 300 answers per column count: one base and one model answer lack a verdict from one judge.)
112
+
113
+ 2 of 3: 104 questions where only the base invents, 13 where only this model does
114
+ (McNemar p < 1e-18).
115
+
116
+ The same judges on the questions where the recipe **is** shown and the model has to read it (290 questions; 2 of 3
117
+ must agree): GLM-4-9B-0414 correct 83.1% [78.4, 87.0], this model **96.2%**
118
+ [93.3, 97.9] (42 questions only this model gets right, 4 only the base
119
+ does, McNemar p < 1e-8). Two other kinds of question in the set are reported apart: 82 general questions graded
120
+ against a reference answer (56.1% base, 46.3% this model: lower, but not significantly, McNemar p = 0.12 with 14
121
+ questions only the base gets right and 6 only this model; see the quantised files below) and
122
+ 10 "how do you make X?" questions with no recipe whose reference is the original 1800s text (20.0% /
123
+ 10.0%: the model is built not to recall historical recipes from memory).
124
+
125
+ **A fourth judge.** Qwen3.8-27B (Apache-2.0) also scored every answer. It is a Qwen-family model and wrote the training
126
+ questions, so it is not independent of the training data and is kept out of the headline. With all four judges and 3
127
+ of 4 agreeing: invents 30.1% (base) vs **4.7%**; correct with the recipe shown
128
+ 78.6% vs **95.5%**.
129
+
130
+ **Next to the Qwen3-8B sibling** (same questions, same three judges, both models' answers generated the same way):
131
+
132
+ | | invents, no recipe (2 of 3) | correct, recipe shown (2 of 3) |
133
+ |---|---|---|
134
+ | Qwen3-8B → Qwen3-8B Cooking Assistant v1 | 43.0% → 10.3% | 85.9% → 95.5% |
135
+ | GLM-4-9B-0414 → this model | 40.1% → 9.7% | 83.1% → 96.2% |
136
+
137
+ (The Qwen card's own figures come from a four-judge panel that included GLM-4.7-Flash, so they differ from this row.)
138
+
139
+ Model judges tend to make the same mistakes as each other, so three agreeing is less than three independent checks.
140
+ On the no-recipe questions the judges agree less about this model's answers than about the base's (Fleiss kappa 0.33
141
+ vs 0.58).
142
+
143
+ **Quantised files** were scored on the same 682 questions against the full-precision model, same judges:
144
+
145
+ | | Q4_K_M | Q6_K | Q8_0 |
146
+ |---|---|---|---|
147
+ | Invents with no recipe, 2 of 3 judges (full precision) | 10.7% (8.7%) | 11.7% (9.3%) | 9.4% (10.0%) |
148
+ | Correct when the recipe is shown, 2 of 3 judges (full precision) | 95.5% (95.9%) | 95.5% (96.2%) | 95.9% (96.2%) |
149
+ | Difference significant? (McNemar, inventing) | no, p = 0.26 | no, p = 0.14 | no, p = 0.77 |
150
+
151
+ With the three headline judges, no quantised file differs significantly from full precision on inventing (p ≥ 0.14) or on reading a shown recipe (p ≥ 0.5). On the 82 general questions every quantised file scored higher than full precision: Q4_K_M 54.9% vs 46.3% (p = 0.039), Q6_K 53.7% vs 47.6% (p = 0.27), Q8_0 56.1% vs 47.6% (p = 0.065). The full-precision answers are the same text in every comparison; only the judging was re-run for each file, which alone moves their scores by up to 1.3 points. The quantised answers come from llama.cpp and the full-precision ones from vLLM, so the full-precision model's lower general score may be partly a serving or decoding effect (one of three differences has p < 0.05, none after correcting for three comparisons); not settled.
152
+
153
+ **Memorisation.** Training documents from the Wikibooks Cookbook (CC BY-SA) were probed with prefixes of 32 to 128 tokens at several offsets, as raw text and as chat prompts, with greedy decoding. The model never continues a document verbatim (no 50-token match; longest overlap 89 characters, the same as base GLM-4-9B-0414), and its overlap with the true text is no higher than the base's (ROUGE-L 0.158 vs 0.161).
154
+
155
+ ## How it was built
156
+
157
+ | | |
158
+ |---|---|
159
+ | Sources | Cooking texts in 26 languages from openly licensed and public-domain sources only: the Wikibooks Cookbook (CC BY-SA 4.0), public-domain cookbooks from Project Gutenberg and the Internet Archive (scans re-read with olmOCR-2), and NHLBI recipes (US government, public domain). Every record carries its source URL, licence and public-domain basis |
160
+ | Excluded | Traditional and Indigenous knowledge, sacred material, outsider-authored books about other cultures, dog-meat recipes, anything without a verified licence |
161
+ | Models used | Open-weights models only, all Apache-2.0 or MIT. **No output from any closed or API model is in the training, evaluation or seed data** |
162
+ | Training | The training data and both stages of Qwen3-8B Cooking Assistant v1, unchanged, on GLM-4-9B-0414: two LoRA stages, each merged: a supervised stage that teaches reading a shown recipe, asking when there is none, and full general answers; then a preference stage that favours asking over inventing and correct over wrong. General chat data is mixed in to limit forgetting |
163
+ | Tooling | LLaMA-Factory, vLLM, llama.cpp |
164
+
165
+ The evaluation set behind every number on this card (all 682 questions, both models' answers, every judge's verdict,
166
+ and per-row source credits) is published so the results can be checked:
167
+ [HumanjavaEnterprises/cooking-assistant-eval](https://huggingface.co/datasets/HumanjavaEnterprises/cooking-assistant-eval),
168
+ config `glm-4-9b-v1`.
169
+
170
+ ## What carried over, and what did not
171
+
172
+ - **The data was written for the Qwen model.** Part of the general answers were rewritten by base Qwen3-8B (for the
173
+ Qwen model, its own words), and the preference stage uses answer pairs sampled from the Qwen model, not from this
174
+ one. Both were kept unchanged so that only the base model differs; for GLM-4-9B-0414 they are another model's words.
175
+ - **The two cooking behaviours carried over.** On inventing and on reading a shown recipe this model shows no significant difference from Qwen3-8B Cooking Assistant v1 under the same judges (McNemar p = 0.86 and 0.80). **General answers did not:** where the Qwen model stayed level with its base, this model's full-precision score fell below GLM-4-9B-0414's (not significant; see Results).
176
+
177
+ ## Limitations
178
+
179
+ - **Not a source of truth for food safety.** It repeats standard guidance (e.g. 74 °C / 165 °F for poultry) but can be
180
+ wrong; check official sources.
181
+ - **It does not know historical recipes by heart.** It was trained on old cookbooks to read and reason about recipes,
182
+ not to recall them. For a specific historical dish, give it the text.
183
+ - The judges are models. Three open judges from three other families agree on the direction of the changes,
184
+ and we keep checking them against people with Label Test, our blind labelling tool. **Help refine it:** label five
185
+ answers at [label-test.humanjava.com](https://label-test.humanjava.com/?s=cook4&k=cook-9ac741a1e387&src=hf-gguf-glm) (no account; we keep only a name or initials and your labels).
186
+ - Still invents on some no-recipe questions: about 1 in 10 when two of three judges must agree (9.7%), nearly 1 in 4 by the strictest single judge (22.7%).
187
+ - **General answers are not improved.** On 82 general questions the full-precision model scored 46.3% vs the base's 56.1% (not significant, p = 0.12); the quantised files scored 54-56% on the same questions (not compared against the base), so the gap may be partly a serving or decoding effect (one of three differences has p < 0.05, none after correcting for three comparisons); not settled.
188
+ - Best in English and Spanish: it invents on 6.9% and 3.8% of the no-recipe questions there, 12.6% across the other
189
+ languages together (2 of 3 judges).
190
+ - The old public-domain cookbooks reflect their era: measures, techniques, some outdated advice.
191
+
192
+ ## Feedback and security
193
+
194
+ - **The model got something wrong** (invented a detail, misread a recipe, unsafe advice): please
195
+ [open a report](https://github.com/HumanjavaEnterprises/glm-4-9b-cooking-assistant/issues/new?template=wrong-answer.yml).
196
+ Real examples are what improve the next version.
197
+ - **Help measure it:** [label five answers](https://label-test.humanjava.com/?s=cook4&k=cook-9ac741a1e387&src=hf-gguf-glm) on Label Test.
198
+ - **Security** (files that don't match `SHA256SUMS`, unexpected embedded content): report privately, see
199
+ [SECURITY.md](https://github.com/HumanjavaEnterprises/glm-4-9b-cooking-assistant/blob/main/SECURITY.md).
200
+
201
+ ## Licence and credits
202
+
203
+ Weights: MIT, same as the base model. Base model: GLM-4-9B-0414 by Zhipu AI / Z.ai (MIT); its copyright and permission
204
+ notice is kept in `LICENSE`:
205
+
206
+ > Copyright (c) 2025 Zhipu AI. Permission is hereby granted, free of charge, to any person obtaining a copy of this
207
+ > software and associated documentation files (the "Software"), to deal in the Software without restriction [...]
208
+ > The above copyright notice and this permission notice shall be included in all copies or substantial portions of
209
+ > the Software. (Full text in `LICENSE`.)
210
+
211
+ Training data sources: Wikibooks Cookbook contributors (CC BY-SA 4.0; per-page attribution in the dataset), Project
212
+ Gutenberg and Internet Archive public-domain books, NHLBI. General chat data: NVIDIA
213
+ (Nemotron-Instruction-Following-Chat-v1, ODC-BY 1.0, responses by GPT-OSS-120B and Qwen3-235B-A22B), Hugging Face
214
+ (SmolTalk2, Apache-2.0, responses by Qwen3-32B; prompts from SystemChat-2.0, the Aya Dataset by Cohere For AI, and
215
+ Table-GPT by Microsoft). Teacher: Qwen3.8-27B (Apache-2.0). Rewritten general answers: Qwen3-8B (Apache-2.0).
216
+ Preference-stage answers: Qwen3-8B Cooking Assistant (Apache-2.0). Judges: DeepSeek-V4-Flash (MIT), gpt-oss-120b
217
+ (Apache-2.0), Seed-OSS-36B-Instruct by ByteDance Seed (Apache-2.0), Qwen3.8-27B (Apache-2.0). OCR: olmOCR-2 by Ai2
218
+ (Apache-2.0). Training: LLaMA-Factory. Quantisation: llama.cpp.
219
+
220
+ Built by [Humanjava Enterprises](https://huggingface.co/HumanjavaEnterprises).
SHA256SUMS ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ 7b8382facf3852614641d35709bfa935a77a678d0d5b8f2f8308658d28805f70 ./glm-4-9b-cooking-assistant-v1-BF16.gguf
2
+ bb34f10b7bb7e59257b3f877538be6581c98f8a436653c063b0652dc3d01d306 ./glm-4-9b-cooking-assistant-v1.imatrix.gguf
3
+ 5cadc80275fa5e4ae4b2d82b123963bcbc20d9c03c00a401b789510b460c6eac ./glm-4-9b-cooking-assistant-v1-Q4_K_M.gguf
4
+ 42bcc552f7d53808abc0641f3dbdf62254ec792cb04d0e02b1c26e650cc0a228 ./glm-4-9b-cooking-assistant-v1-Q6_K.gguf
5
+ 966b7413b1fc1906c45f98a140e6cd55cbf814e586ad019c21bcf218362736ff ./glm-4-9b-cooking-assistant-v1-Q8_0.gguf
glm-4-9b-cooking-assistant-v1-BF16.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b8382facf3852614641d35709bfa935a77a678d0d5b8f2f8308658d28805f70
3
+ size 18811581536
glm-4-9b-cooking-assistant-v1-Q4_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5cadc80275fa5e4ae4b2d82b123963bcbc20d9c03c00a401b789510b460c6eac
3
+ size 6166574400
glm-4-9b-cooking-assistant-v1-Q6_K.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:42bcc552f7d53808abc0641f3dbdf62254ec792cb04d0e02b1c26e650cc0a228
3
+ size 8266642752
glm-4-9b-cooking-assistant-v1-Q8_0.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:966b7413b1fc1906c45f98a140e6cd55cbf814e586ad019c21bcf218362736ff
3
+ size 9999610976
glm-4-9b-cooking-assistant-v1.imatrix.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb34f10b7bb7e59257b3f877538be6581c98f8a436653c063b0652dc3d01d306
3
+ size 11017312
params ADDED
@@ -0,0 +1 @@
 
 
1
+ {"stop": ["<|user|>", "<|observation|>", "<|endoftext|>"], "temperature": 0.7, "top_p": 0.8, "top_k": 20, "num_ctx": 8192}
system ADDED
@@ -0,0 +1 @@
 
 
1
+ You are a cooking assistant. You answer questions about dishes, ingredients, techniques and food safety. When you are not sure, you say so.
template ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ [gMASK]<sop>
2
+ {{- range .Messages }}
3
+ {{- if eq .Role "system" }}<|system|>
4
+ {{ .Content }}
5
+ {{- else if eq .Role "user" }}<|user|>
6
+ {{ .Content }}
7
+ {{- else if eq .Role "assistant" }}<|assistant|>
8
+ {{ .Content }}
9
+ {{- end }}
10
+ {{- end }}<|assistant|>