gyung commited on
Commit
d5e64d3
ยท
verified ยท
1 Parent(s): f494ce3

Update model card with representative finance QA smoke test

Browse files
README.md CHANGED
@@ -62,7 +62,7 @@ Download a quantized GGUF file:
62
 
63
  ```bash
64
  huggingface-cli download LLM-OS-Models/KoHRM-Text-1.4B-GGUF \
65
- KoHRM-Text-1.4B-Q4_K_M.gguf \
66
  --local-dir .
67
  ```
68
 
@@ -70,57 +70,70 @@ Run on CPU:
70
 
71
  ```bash
72
  ./build-hrm/bin/llama-cli \
73
- -m ./KoHRM-Text-1.4B-Q4_K_M.gguf \
74
  -ngl 0 \
75
  -t 4 \
76
- -c 512 \
77
- -n 80 \
78
- --seed 42 \
79
- --temp 0.7 \
80
  --top-p 0.9 \
 
81
  --single-turn \
82
  --simple-io \
83
  --no-warmup \
84
  --display-prompt \
85
- -p $'์งˆ๋ฌธ: ๋Œ€ํ•œ๋ฏผ๊ตญ ๋ฏผ๋ฒ•์ƒ ๋ฏธ์„ฑ๋…„์ž์˜ ๋ฒ•๋ฅ ํ–‰์œ„ ์ทจ์†Œ์— ๋Œ€ํ•ด ํ•œ ๋ฌธ๋‹จ์œผ๋กœ ์„ค๋ช…ํ•˜๋ผ.\n๋‹ต๋ณ€:'
86
  ```
87
 
88
  ## CPU Generation Tests
89
 
90
  Tested locally on CPU with the patched `llama.cpp` build and the prompt shown above.
91
 
92
- | file | prompt speed | generation speed | value check |
93
- |---|---:|---:|---|
94
- | `KoHRM-Text-1.4B-Q4_K_M.gguf` | 33.6 t/s | 13.2 t/s | Runtime OK, not suitable as a legal-answer quality example |
95
- | `KoHRM-Text-1.4B-Q8_0.gguf` | 23.8 t/s | 6.8 t/s | Runtime OK, not suitable as a legal-answer quality example |
96
-
97
- Q4_K_M output excerpt:
98
 
99
  ```text
100
- The user wants a single-word, concise explanation ... The core idea is:
101
- ๋ฏธ์„ฑ๋…„์ž๋Š” ๋ฏธ์„ฑ๋…„์ž๋ฅผ ๋Œ€์ƒ์œผ๋กœ ํ•˜๋Š” ๋ฒ•๋ฅ ํ–‰์œ„ ... ๋ฏผ๋ฒ• ์ œ999-1001 ...
102
  ```
103
 
 
 
 
 
 
 
 
 
 
104
  Q8_0 output excerpt:
105
 
106
  ```text
107
- The user wants a single-word, concise explanation ... The key insight is
108
- to use "์†Œ์œ„ '๋ฏธ์„ฑ๋…„์ž'" ...
 
 
 
 
109
  ```
110
 
111
- These are valuable as reproducible CPU runtime checks, but not as meaningful legal QA outputs. The outputs include meta-instruction text and legally incorrect or incoherent content. This matches the source checkpoint status: the model is a rolling pretraining-stage checkpoint, not a final instruction-tuned chat/legal model.
112
-
113
- The full smoke-test logs are in `reports/`.
114
 
115
  ## Prompt Format
116
 
117
- The source model card documents this project-side prompt wrapper:
118
 
119
  ```text
120
  <|im_start|><|object_ref_start|>PROMPT<|im_end|>
121
  ```
122
 
123
- The public checkpoint is a rolling pretraining-stage model, not a final chat/SFT model. Expect unstable instruction-following behavior and possible immediate end-token generation depending on prompt and sampler settings. The GGUF files here validate conversion and patched CPU runtime support, not final assistant quality.
 
 
 
 
 
124
 
125
  ## GGUF Metadata
126
 
 
62
 
63
  ```bash
64
  huggingface-cli download LLM-OS-Models/KoHRM-Text-1.4B-GGUF \
65
+ KoHRM-Text-1.4B-Q8_0.gguf \
66
  --local-dir .
67
  ```
68
 
 
70
 
71
  ```bash
72
  ./build-hrm/bin/llama-cli \
73
+ -m ./KoHRM-Text-1.4B-Q8_0.gguf \
74
  -ngl 0 \
75
  -t 4 \
76
+ -c 1024 \
77
+ -n 260 \
78
+ --seed 41 \
79
+ --temp 0.45 \
80
  --top-p 0.9 \
81
+ --repeat-penalty 1.08 \
82
  --single-turn \
83
  --simple-io \
84
  --no-warmup \
85
  --display-prompt \
86
+ -p $'ํ•ด์™ธ์ฃผ์‹ ํˆฌ์ž์—์„œ ์›/๋‹ฌ๋Ÿฌ ํ™˜์œจ ๋ณ€๋™์ด ์›ํ™” ์ˆ˜์ต๋ฅ ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ๋Œ€์‘ ๋ฐฉ์•ˆ์„ ๊ฐ„๋‹จํžˆ ์„ค๋ช…ํ•ด ์ฃผ์„ธ์š”.'
87
  ```
88
 
89
  ## CPU Generation Tests
90
 
91
  Tested locally on CPU with the patched `llama.cpp` build and the prompt shown above.
92
 
93
+ This prompt was chosen after checking the KoHRM training-data path. KoHRM uses
94
+ the HRM `V1Dataset` instruction-response layout:
 
 
 
 
95
 
96
  ```text
97
+ <|im_start|><condition_token>instruction<|im_end|>response<|box_end|>
 
98
  ```
99
 
100
+ The instruction/prefix span is not trained with loss; the response span is
101
+ trained with response-only loss. Local decoded samples include short Korean
102
+ finance QA rows, so the representative GGUF smoke prompt below uses the same
103
+ plain instruction style instead of a legal reasoning prompt.
104
+
105
+ | file | prompt speed | generation speed | value check |
106
+ |---|---:|---:|---|
107
+ | `KoHRM-Text-1.4B-Q8_0.gguf` | 25.3 t/s | 5.0 t/s | Runtime OK; useful qualitative finance QA sample |
108
+
109
  Q8_0 output excerpt:
110
 
111
  ```text
112
+ 1. **ํ™˜์œจ ๋ณ€๋™์˜ ์˜ํ–ฅ:** ํ•ด์™ธ ์ฃผ์‹ ํˆฌ์ž์˜ ์ˆ˜์ต๋ฅ ์€ ์ฃผ๊ฐ€ ์ƒ์Šน์— ๋”ฐ๋ฅธ ์ˆ˜์ต๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์ˆ˜์ต ๋˜๋Š” ์†์‹ค์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
113
+ ...
114
+ 2. **๋Œ€์‘ ๋ฐฉ์•ˆ:**
115
+ - **ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜ ๊ด€๋ฆฌ:** ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ํ™˜ํ—ค์ง€ ์ƒํ’ˆ์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜, ๋ถ„ํ•  ๋งค์ˆ˜/๋งค๋„ ์ „๋žต์„ ํ†ตํ•ด ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์˜ํ–ฅ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
116
+ - **์žฅ๊ธฐ ํˆฌ์ž:** ์žฅ๊ธฐ ํˆฌ์ž๋ฅผ ํ†ตํ•ด ํ™˜์œจ ๋ณ€๋™์˜ ๋‹จ๊ธฐ์ ์ธ ์˜ํ–ฅ์„ ์™„ํ™”ํ•˜๊ณ , ์žฅ๊ธฐ์ ์ธ ์ฃผ๊ฐ€ ์ƒ์Šน์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
117
+ - **๋ถ„์‚ฐ ํˆฌ์ž:** ๋‹ค์–‘ํ•œ ๊ตญ๊ฐ€์˜ ์ฃผ์‹์— ๋ถ„์‚ฐ ํˆฌ์žํ•˜์—ฌ ํŠน์ • ๊ตญ๊ฐ€์˜ ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
118
  ```
119
 
120
+ The full smoke-test log is in `reports/generation_tests/finance_short_q8_0.txt`.
121
+ This is a qualitative CPU runtime sample, not a benchmark or financial advice.
 
122
 
123
  ## Prompt Format
124
 
125
+ The source training/inference wrapper is:
126
 
127
  ```text
128
  <|im_start|><|object_ref_start|>PROMPT<|im_end|>
129
  ```
130
 
131
+ `prepare_sft_data.py` writes the generic HRM V1Dataset layout with
132
+ `direct=<|object_ref_start|>` by default. In this patched GGUF runtime,
133
+ `llama-completion` could load the model but returned an immediate end token for
134
+ the tested prompts, while `llama-cli --single-turn` produced visible CPU token
135
+ generation. The public checkpoint is a rolling pretraining-stage model, not a
136
+ final chat/SFT model, so instruction following can still be unstable.
137
 
138
  ## GGUF Metadata
139
 
reports/conversion_report.md CHANGED
@@ -93,11 +93,16 @@ Exiting...
93
 
94
  The generated text is not a legal-quality answer; this source checkpoint is documented as a pretraining-stage model, not an instruction-tuned chat model.
95
 
96
- Additional CPU generation tests were run on 2026-06-14 with Q4_K_M and Q8_0 using the model-card prompt. Both loaded and generated on CPU, but both were rejected as meaningful legal QA examples because they contained meta-instruction text and legally incorrect or incoherent content.
 
 
 
97
 
98
  ```text
99
- Q4_K_M: Prompt 33.6 t/s, Generation 13.2 t/s, runtime OK, quality example rejected
100
- Q8_0: Prompt 23.8 t/s, Generation 6.8 t/s, runtime OK, quality example rejected
 
 
101
  ```
102
 
103
  ## Artifacts
 
93
 
94
  The generated text is not a legal-quality answer; this source checkpoint is documented as a pretraining-stage model, not an instruction-tuned chat model.
95
 
96
+ Additional CPU generation tests were run on 2026-06-14. After checking the KoHRM
97
+ training-data layout, the public model-card example was changed from a legal QA
98
+ prompt to a short Korean finance QA prompt that better matches the local decoded
99
+ prepared samples and the instruction-response preprocessing path.
100
 
101
  ```text
102
+ Q8_0 prompt:
103
+ ํ•ด์™ธ์ฃผ์‹ ํˆฌ์ž์—์„œ ์›/๋‹ฌ๋Ÿฌ ํ™˜์œจ ๋ณ€๋™์ด ์›ํ™” ์ˆ˜์ต๋ฅ ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ๋Œ€์‘ ๋ฐฉ์•ˆ์„ ๊ฐ„๋‹จํžˆ ์„ค๋ช…ํ•ด ์ฃผ์„ธ์š”.
104
+
105
+ Q8_0: Prompt 25.3 t/s, Generation 5.0 t/s, runtime OK, useful qualitative finance QA sample
106
  ```
107
 
108
  ## Artifacts
reports/generation_quality_check.md CHANGED
@@ -2,48 +2,82 @@
2
 
3
  Date: 2026-06-14 UTC
4
 
5
- Prompt from the model card:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
 
7
  ```text
8
- ์งˆ๋ฌธ: ๋Œ€ํ•œ๋ฏผ๊ตญ ๋ฏผ๋ฒ•์ƒ ๋ฏธ์„ฑ๋…„์ž์˜ ๋ฒ•๋ฅ ํ–‰์œ„ ์ทจ์†Œ์— ๋Œ€ํ•ด ํ•œ ๋ฌธ๋‹จ์œผ๋กœ ์„ค๋ช…ํ•˜๋ผ.
9
- ๋‹ต๋ณ€:
10
  ```
11
 
12
- Command pattern:
13
 
14
  ```bash
15
  ./build-hrm/bin/llama-cli \
16
- -m MODEL.gguf \
17
  -ngl 0 \
18
  -t 4 \
19
- -c 512 \
20
- -n 80 \
21
- --seed 42 \
22
- --temp 0.7 \
23
  --top-p 0.9 \
 
24
  --single-turn \
25
  --simple-io \
26
  --no-warmup \
27
  --display-prompt \
28
- -p $'์งˆ๋ฌธ: ๋Œ€ํ•œ๋ฏผ๊ตญ ๋ฏผ๋ฒ•์ƒ ๋ฏธ์„ฑ๋…„์ž์˜ ๋ฒ•๋ฅ ํ–‰์œ„ ์ทจ์†Œ์— ๋Œ€ํ•ด ํ•œ ๋ฌธ๋‹จ์œผ๋กœ ์„ค๋ช…ํ•˜๋ผ.\n๋‹ต๋ณ€:'
29
  ```
30
 
31
- ## Results
32
 
33
  | model | prompt speed | generation speed | verdict |
34
  |---|---:|---:|---|
35
- | `KoHRM-Text-1.4B-Q4_K_M.gguf` | 33.6 t/s | 13.2 t/s | Runtime OK, not suitable as a legal-answer quality example |
36
- | `KoHRM-Text-1.4B-Q8_0.gguf` | 23.8 t/s | 6.8 t/s | Runtime OK, not suitable as a legal-answer quality example |
 
37
 
38
- ## Value Assessment
 
 
 
 
 
 
 
39
 
40
- The outputs prove that the patched `llama.cpp` runtime can load and generate from both Q4_K_M and Q8_0 GGUF files on CPU.
 
 
 
 
41
 
42
- The outputs should not be presented as useful legal QA examples:
43
 
44
- - They contain meta-instruction text such as `The user wants...`.
45
- - They include legally incorrect or incoherent claims.
46
- - The Q4_K_M sample references `๋ฏผ๋ฒ• ์ œ999-1001` for ๋ฏธ์„ฑ๋…„์ž ํ–‰์œ„ ์ œํ•œ, which is not a valid basis for the asked civil-law issue.
47
- - The Q8_0 sample is semantically incoherent around `์†Œ์œ„ '๋ฏธ์„ฑ๋…„์ž'`, `์ฒ˜๋ถ„`, and `์˜๋ฌด`.
48
 
49
- Conclusion: the generation results are valuable only as reproducible runtime smoke tests. They are not evidence of instruction-following quality or legal reasoning quality. This is consistent with the source model being a pretraining-stage checkpoint rather than an SFT/chat/legal model.
 
 
2
 
3
  Date: 2026-06-14 UTC
4
 
5
+ ## Data Format Check
6
+
7
+ Before selecting the representative generation example, the local KoHRM data
8
+ pipeline was checked from the repository docs and code.
9
+
10
+ KoHRM uses the HRM `V1Dataset` instruction-response layout:
11
+
12
+ ```text
13
+ <|im_start|><condition_token>instruction<|im_end|>response<|box_end|>
14
+ ```
15
+
16
+ Important preprocessing details:
17
+
18
+ - `scripts/prepare_sft_data.py` accepts rows with `instruction`, `response`, and
19
+ optional `condition`.
20
+ - The default direct condition is `direct=<|object_ref_start|>`.
21
+ - `dataset_new.py` reads instruction and response spans separately.
22
+ - With `target_only=True`, the instruction/prefix span has ignore labels and
23
+ the response span receives the supervised loss.
24
+ - Local decoded prepared samples include short Korean finance QA examples, so a
25
+ finance QA prompt is a better smoke-test example than a legal reasoning prompt.
26
+
27
+ ## Representative Prompt
28
 
29
  ```text
30
+ ํ•ด์™ธ์ฃผ์‹ ํˆฌ์ž์—์„œ ์›/๋‹ฌ๋Ÿฌ ํ™˜์œจ ๋ณ€๋™์ด ์›ํ™” ์ˆ˜์ต๋ฅ ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ๋Œ€์‘ ๋ฐฉ์•ˆ์„ ๊ฐ„๋‹จํžˆ ์„ค๋ช…ํ•ด ์ฃผ์„ธ์š”.
 
31
  ```
32
 
33
+ Command:
34
 
35
  ```bash
36
  ./build-hrm/bin/llama-cli \
37
+ -m KoHRM-Text-1.4B-Q8_0.gguf \
38
  -ngl 0 \
39
  -t 4 \
40
+ -c 1024 \
41
+ -n 260 \
42
+ --seed 41 \
43
+ --temp 0.45 \
44
  --top-p 0.9 \
45
+ --repeat-penalty 1.08 \
46
  --single-turn \
47
  --simple-io \
48
  --no-warmup \
49
  --display-prompt \
50
+ -p $'ํ•ด์™ธ์ฃผ์‹ ํˆฌ์ž์—์„œ ์›/๋‹ฌ๋Ÿฌ ํ™˜์œจ ๋ณ€๋™์ด ์›ํ™” ์ˆ˜์ต๋ฅ ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ๋Œ€์‘ ๋ฐฉ์•ˆ์„ ๊ฐ„๋‹จํžˆ ์„ค๋ช…ํ•ด ์ฃผ์„ธ์š”.'
51
  ```
52
 
53
+ ## Result
54
 
55
  | model | prompt speed | generation speed | verdict |
56
  |---|---:|---:|---|
57
+ | `KoHRM-Text-1.4B-Q8_0.gguf` | 25.3 t/s | 5.0 t/s | Runtime OK; useful qualitative finance QA sample |
58
+
59
+ Excerpt:
60
 
61
+ ```text
62
+ 1. **ํ™˜์œจ ๋ณ€๋™์˜ ์˜ํ–ฅ:** ํ•ด์™ธ ์ฃผ์‹ ํˆฌ์ž์˜ ์ˆ˜์ต๋ฅ ์€ ์ฃผ๊ฐ€ ์ƒ์Šน์— ๋”ฐ๋ฅธ ์ˆ˜์ต๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์ˆ˜์ต ๋˜๋Š” ์†์‹ค์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
63
+ ...
64
+ 2. **๋Œ€์‘ ๋ฐฉ์•ˆ:**
65
+ - **ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜ ๊ด€๋ฆฌ:** ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ํ™˜ํ—ค์ง€ ์ƒํ’ˆ์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜, ๋ถ„ํ•  ๋งค์ˆ˜/๋งค๋„ ์ „๋žต์„ ํ†ตํ•ด ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์˜ํ–ฅ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
66
+ - **์žฅ๊ธฐ ํˆฌ์ž:** ์žฅ๊ธฐ ํˆฌ์ž๋ฅผ ํ†ตํ•ด ํ™˜์œจ ๋ณ€๋™์˜ ๋‹จ๊ธฐ์ ์ธ ์˜ํ–ฅ์„ ์™„ํ™”ํ•˜๊ณ , ์žฅ๊ธฐ์ ์ธ ์ฃผ๊ฐ€ ์ƒ์Šน์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
67
+ - **๋ถ„์‚ฐ ํˆฌ์ž:** ๋‹ค์–‘ํ•œ ๊ตญ๊ฐ€์˜ ์ฃผ์‹์— ๋ถ„์‚ฐ ํˆฌ์žํ•˜์—ฌ ํŠน์ • ๊ตญ๊ฐ€์˜ ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
68
+ ```
69
 
70
+ Full log:
71
+
72
+ ```text
73
+ reports/generation_tests/finance_short_q8_0.txt
74
+ ```
75
 
76
+ ## Notes
77
 
78
+ `llama-completion` loaded the patched GGUF but returned an immediate end token
79
+ for tested prompts. `llama-cli --single-turn` produced visible CPU token
80
+ generation and is therefore used for the model-card smoke command.
 
81
 
82
+ This is a qualitative CPU runtime check, not a benchmark, not financial advice,
83
+ and not evidence that the pretraining-stage checkpoint is a final chat/SFT model.
reports/generation_tests/finance_short_q8_0.txt ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ warning: no usable GPU found, --gpu-layers option will be ignored
2
+ warning: one possible reason is that llama.cpp was compiled without GPU support
3
+ warning: consult docs/build.md for compilation instructions
4
+
5
+ Loading model...
6
+
7
+
8
+ โ–„โ–„ โ–„โ–„
9
+ โ–ˆโ–ˆ โ–ˆโ–ˆ
10
+ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–€โ–€โ–ˆโ–„ โ–ˆโ–ˆโ–ˆโ–„โ–ˆโ–ˆโ–ˆโ–„ โ–€โ–€โ–ˆโ–„ โ–„โ–ˆโ–ˆโ–ˆโ–ˆ โ–ˆโ–ˆโ–ˆโ–ˆโ–„ โ–ˆโ–ˆโ–ˆโ–ˆโ–„
11
+ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–„โ–ˆโ–€โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–„โ–ˆโ–€โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ
12
+ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–€โ–ˆโ–„โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–ˆโ–ˆ โ–€โ–ˆโ–„โ–ˆโ–ˆ โ–ˆโ–ˆ โ–€โ–ˆโ–ˆโ–ˆโ–ˆ โ–ˆโ–ˆโ–ˆโ–ˆโ–€ โ–ˆโ–ˆโ–ˆโ–ˆโ–€
13
+ โ–ˆโ–ˆ โ–ˆโ–ˆ
14
+ โ–€โ–€ โ–€โ–€
15
+
16
+ build : b9263-6a257d446
17
+ model : KoHRM-Text-1.4B-Q8_0.gguf
18
+ modalities : text
19
+
20
+ available commands:
21
+ /exit or Ctrl+C stop or exit
22
+ /regen regenerate the last response
23
+ /clear clear the chat history
24
+ /read <file> add a text file
25
+ /glob <pattern> add text files using globbing pattern
26
+
27
+
28
+ > ํ•ด์™ธ์ฃผ์‹ ํˆฌ์ž์—์„œ ์›/๋‹ฌ๋Ÿฌ ํ™˜์œจ ๋ณ€๋™์ด ์›ํ™” ์ˆ˜์ต๋ฅ ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ๋Œ€์‘ ๋ฐฉ์•ˆ์„ ๊ฐ„๋‹จํžˆ ์„ค๋ช…ํ•ด ์ฃผ์„ธ์š”.
29
+
30
+
31
+ 1. **ํ™˜์œจ ๋ณ€๋™์˜ ์˜ํ–ฅ:** ํ•ด์™ธ ์ฃผ์‹ ํˆฌ์ž์˜ ์ˆ˜์ต๋ฅ ์€ ์ฃผ๊ฐ€ ์ƒ์Šน์— ๋”ฐ๋ฅธ ์ˆ˜์ต๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์ˆ˜์ต ๋˜๋Š” ์†์‹ค์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฏธ๊ตญ ์ฃผ์‹์„ ๋งค์ˆ˜ํ–ˆ์„ ๋•Œ ํ™˜์œจ์ด 1,200์›์ด์—ˆ๊ณ , ๋งค๋„ ์‹œ์ ์— ํ™˜์œจ์ด 1,300์›์œผ๋กœ ์ƒ์Šนํ–ˆ๋‹ค๋ฉด, ์ฃผ๊ฐ€ ์ƒ์Šน๊ณผ ํ•จ๊ป˜ ํ™˜์ฐจ์ต์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ํ™˜์œจ์ด ํ•˜๋ฝํ–ˆ๋‹ค๋ฉด ํ™˜์ฐจ์†์ด ๋ฐœ์ƒํ•˜์—ฌ ์ „์ฒด ์ˆ˜์ต๋ฅ ์ด ๋‚ฎ์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
32
+
33
+ 2. **๋Œ€์‘ ๋ฐฉ์•ˆ:**
34
+ - **ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜ ๊ด€๋ฆฌ:** ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ํ™˜ํ—ค์ง€ ์ƒํ’ˆ์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜, ๋ถ„ํ•  ๋งค์ˆ˜/๋งค๋„ ์ „๋žต์„ ํ†ตํ•ด ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์˜ํ–ฅ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
35
+ - **์žฅ๊ธฐ ํˆฌ์ž:** ์žฅ๊ธฐ ํˆฌ์ž๋ฅผ ํ†ตํ•ด ํ™˜์œจ ๋ณ€๋™์˜ ๋‹จ๊ธฐ์ ์ธ ์˜ํ–ฅ์„ ์™„ํ™”ํ•˜๊ณ , ์žฅ๊ธฐ์ ์ธ ์ฃผ๊ฐ€ ์ƒ์Šน์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
36
+ - **๋ถ„์‚ฐ ํˆฌ์ž:** ๋‹ค์–‘ํ•œ ๊ตญ๊ฐ€์˜ ์ฃผ์‹์— ๋ถ„์‚ฐ ํˆฌ์žํ•˜์—ฌ ํŠน์ • ๊ตญ๊ฐ€์˜ ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
37
+ - **ํ™˜์œจ ์ „๋ง:** ํ™˜์œจ ์ „๋ง์„ ์ฐธ๊ณ ํ•˜์—ฌ ํˆฌ์ž ์‹œ์ ์„ ์กฐ์ ˆํ•˜๊ฑฐ๋‚˜, ํ™˜์œจ ๋ณ€๋™์— ๋”ฐ๋ฅธ ์†์‹ค์„ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ „๋žต์„ ์ˆ˜๋ฆฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
38
+
39
+ 3. **์ถ”๊ฐ€ ๊ณ ๋ ค ์‚ฌํ•ญ:**
40
+ - **ํˆฌ์ž ๋ชฉํ‘œ:** ํˆฌ์ž ๋ชฉํ‘œ์— ๋”ฐ๋ผ ์ ์ ˆํ•œ ๋Œ€์‘ ๋ฐฉ์•ˆ์„ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
41
+ - **์œ„ํ—˜ ๊ฐ์ˆ˜ ์ˆ˜์ค€:** ์œ„ํ—˜ ๊ฐ์ˆ˜ ์ˆ˜์ค€์— ๋”ฐ๋ผ ํ™˜์œจ ๋ณ€๋™ ์œ„ํ—˜์„ ๊ด€๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ์ •๋„๋ฅผ ๊ฒฐ์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
42
+ - **์ „๋ฌธ๊ฐ€ ์ž๋ฌธ:** ํ•„์š”์— ๋”ฐ๋ผ ๊ธˆ์œต ์ „๋ฌธ๊ฐ€์˜ ์ž๋ฌธ์„ ๋ฐ›์•„
43
+
44
+ [ Prompt: 25.3 t/s | Generation: 5.0 t/s ]
45
+
46
+ Exiting...