File size: 23,795 Bytes
f5d6a7c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
🦥 Unsloth Zoo will now patch everything to make training faster!
Traceback (most recent call last):
  File "/home/priya/qwen_unsloth/train_qwen38-3S.py", line 53, in <module>
    model, tokenizer = FastModel.from_pretrained(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/priya/miniconda3/envs/geocoding/lib/python3.11/site-packages/unsloth/models/loader_utils.py", line 1278, in _wrapper
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/home/priya/miniconda3/envs/geocoding/lib/python3.11/site-packages/unsloth/models/loader.py", line 1360, in from_pretrained
    new_model_name = get_model_name(
                     ^^^^^^^^^^^^^^^
  File "/home/priya/miniconda3/envs/geocoding/lib/python3.11/site-packages/unsloth/models/loader_utils.py", line 301, in get_model_name
    raise NotImplementedError(
NotImplementedError: Unsloth: unsloth/Qwen3.8-27B-unsloth-bnb-4bit is not supported in your current Unsloth version! Please update Unsloth via:

pip uninstall unsloth unsloth_zoo -y
pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install --upgrade --no-cache-dir "git+https://github.com/unslothai/unsloth-zoo.git"


Starting with the exact official 4-bit model repository.
🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
🦥 Unsloth Zoo will now patch everything to make training faster!
[unsloth_zoo.log|WARNING]Loading unsloth/Qwen3.8-27B-unsloth-bnb-4bit as text-only; vision/audio towers skipped. Use FastVisionModel for multimodal inputs.
==((====))==  Unsloth 2026.8.18: Fast Qwen3_5 patching. Transformers: 5.5.0.
   \\   /|    NVIDIA RTX A6000. Num GPUs = 1. Max memory: 47.527 GB. Platform: Linux.
O^O/ \_/ \    Torch: 2.7.1+cu118. CUDA: 8.6. CUDA Toolkit: 11.8. Triton: 3.3.1
\        /    Bfloat16 = TRUE. FA [Xformers = 0.0.31. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!
The fast path is not available because one of the required library is not installed. Falling back to torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and https://github.com/Dao-AILab/causal-conv1d

Loading weights:   0%|          | 0/851 [00:00<?, ?it/s]
Loading weights:   0%|          | 1/851 [00:10<2:30:59, 10.66s/it]
Loading weights: 100%|██████████| 851/851 [00:10<00:00, 79.17it/s]
Unsloth: Dropout = 0 is supported for fast patching. You are using dropout = 0.1.
Unsloth will patch all other layers, except LoRA matrices, causing a performance hit.

Generating train split: 0 examples [00:00, ? examples/s]
Generating train split: 10000 examples [00:00, 12416.07 examples/s]
Generating train split: 20000 examples [00:01, 16829.55 examples/s]
Generating train split: 29153 examples [00:01, 19020.78 examples/s]
Generating train split: 29153 examples [00:01, 17637.35 examples/s]

Filter:   0%|          | 0/29153 [00:00<?, ? examples/s]
Filter:  21%|██        | 6000/29153 [00:00<00:00, 49617.45 examples/s]
Filter:  69%|██████▊   | 20000/29153 [00:00<00:00, 94226.93 examples/s]
Filter: 100%|██████████| 29153/29153 [00:00<00:00, 96875.34 examples/s]

Filter:   0%|          | 0/29153 [00:00<?, ? examples/s]
Filter:  48%|████▊     | 14000/29153 [00:00<00:00, 130377.62 examples/s]
Filter:  96%|█████████▌| 28000/29153 [00:00<00:00, 134025.06 examples/s]
Filter: 100%|██████████| 29153/29153 [00:00<00:00, 124649.75 examples/s]
Train: 23316
Val: 2918
Train distribution:
Anuṣṭubh: 21763
Indravajrā: 171
Mālinī: 72
Sragdharā: 163
Upendravajrā: 96
Vasantatilakā: 526
Vaṃśastha: 136
Śikhariṇī: 128
Śālinī: 52
Śārdūlavikrīḍita: 209

Validation distribution:
Anuṣṭubh: 2721
Indravajrā: 22
Mālinī: 9
Sragdharā: 20
Upendravajrā: 13
Vasantatilakā: 66
Vaṃśastha: 17
Śikhariṇī: 17
Śālinī: 7
Śārdūlavikrīḍita: 26

Map:   0%|          | 0/23316 [00:00<?, ? examples/s]
Map:   4%|▍         | 1000/23316 [00:00<00:05, 4060.64 examples/s]
Map:   9%|▊         | 2000/23316 [00:00<00:05, 4084.43 examples/s]
Map:  13%|█▎        | 3000/23316 [00:00<00:04, 4200.74 examples/s]
Map:  17%|█▋        | 4000/23316 [00:00<00:04, 4263.75 examples/s]
Map:  21%|██▏       | 5000/23316 [00:01<00:04, 4259.63 examples/s]
Map:  26%|██▌       | 6000/23316 [00:01<00:04, 4212.76 examples/s]
Map:  30%|███       | 7000/23316 [00:01<00:03, 4235.19 examples/s]
Map:  34%|███▍      | 8000/23316 [00:01<00:03, 4248.81 examples/s]
Map:  39%|███▊      | 9000/23316 [00:02<00:03, 4291.32 examples/s]
Map:  43%|████▎     | 10000/23316 [00:02<00:03, 4315.40 examples/s]
Map:  47%|████▋     | 11000/23316 [00:02<00:02, 4166.44 examples/s]
Map:  51%|█████▏    | 12000/23316 [00:02<00:02, 4229.17 examples/s]
Map:  56%|█████▌    | 13000/23316 [00:03<00:02, 4263.01 examples/s]
Map:  60%|██████    | 14000/23316 [00:03<00:02, 4237.17 examples/s]
Map:  64%|██████▍   | 15000/23316 [00:03<00:01, 4274.22 examples/s]
Map:  69%|██████▊   | 16000/23316 [00:03<00:01, 4227.91 examples/s]
Map:  73%|███████▎  | 17000/23316 [00:04<00:01, 4275.01 examples/s]
Map:  77%|███████▋  | 18000/23316 [00:04<00:01, 4295.42 examples/s]
Map:  81%|████████▏ | 19000/23316 [00:04<00:00, 4330.61 examples/s]
Map:  86%|████████▌ | 20000/23316 [00:04<00:00, 4329.39 examples/s]
Map:  90%|█████████ | 21000/23316 [00:04<00:00, 4315.36 examples/s]
Map:  94%|█████████▍| 22000/23316 [00:05<00:00, 4346.47 examples/s]
Map:  99%|█████████▊| 23000/23316 [00:05<00:00, 4357.81 examples/s]
Map: 100%|██████████| 23316/23316 [00:06<00:00, 3645.84 examples/s]

Map:   0%|          | 0/2918 [00:00<?, ? examples/s]
Map:  34%|███▍      | 1000/2918 [00:00<00:00, 4512.13 examples/s]
Map:  69%|██████▊   | 2000/2918 [00:00<00:00, 4346.12 examples/s]
Map: 100%|██████████| 2918/2918 [00:00<00:00, 4298.29 examples/s]
Map: 100%|██████████| 2918/2918 [00:00<00:00, 3667.77 examples/s]
--------------------------------------------------------------------------------------
sa
अहोरात्रेक्षणो दिव्यो वेदाड़ः श्रुतिभूषण: आज्यनास: स्त्रुवातुण्डः सामघोषस्वनो महान्‌॥
--------------------------------------------------------------------------------------

hi
दिन और रात्रि उनके नेत्र हैं, वे दिव्यस्वरूप हैं वेद उनका अड्ग और श्रुतियाँ आभूषण हैं हविष्य नासिका, खुवा थूथन और सामवेदका गम्भीर घोष ही उनका स्वर है वे महान्‌ हैं ॥
--------------------------------------------------------------------------------------

skr_out
anuṣṭubh (1,2: pathyā; 3,4: pathyā)
--------------------------------------------------------------------------------------

syllable_count
32
--------------------------------------------------------------------------------------

skr_meter
Anuṣṭubh
--------------------------------------------------------------------------------------

meter_cd
Anuṣṭubh
--------------------------------------------------------------------------------------

comments
None
--------------------------------------------------------------------------------------

prompt
Objective:

Generate a sanskrit verse in Anuṣṭubh meter based on a given meaning or theme while strictly following the metrical (?and rhyme) rules.

Rules of Chandas:
Syllable Rules:
LAGHU vowels: अ, इ, उ, ऋ, ऌ
GURU vowels: आ, ई, ऊ, ॠ, ॡ, ए, ऐ, ओ, औ
HRASVA vowels: अ, इ, उ, ऋ, ऌ
DEERGHA vowels: आ, ई, ऊ, ॠ, ॡ, ए, ऐ, ओ, औ
Syllable is marked laghu, guru and hrasva, deergha based on the vowel it contains.
Syllable containing anusvāra("ं") or visarga("ः") is always marked as guru.
Syllable that is followed by a conjunct consonant (saṁyuktākṣara)
is always marked guru.

The goal is to generate Sanskrit verse that follows the anushtup meter rules for the given input translation.
RULES:
Verse Rules:
The verse contains 32 syllables/akshara and 4 padas in total.
The verse is divided in 2 lines, each containing 16 syllables.
Each line is divided into 2 padas (quartets),each containing exactly 8 syllables.
The fifth syllable of every pada must be LAGHU or short.
The sixth syllable of every pada must be GURU or long.
The seventh syllable of the second and the fourth pada must be HRASVA
The seventh syllable of the first and third paada must be DEERGHA

Language:
The generated verse must be in sanskrit in Devanagri script.
The Input given is in the hindi.
Do NOT include English words, transliteration, labels, or explanations.

Output Format:
(Output exactly 4 lines, one pada per line. Nothing else.)

Example 1:
Meaning:
इसलिये कामनाओं या भोगोंकी वृद्धिके लिये आग्रह नहीं रखना चाहिये भोगों में जो आसक्ति होती है, वह दुःखरूप ही है धन पाकर भी उसे धर्ममें ही लगा देना चाहिये काम भोगोंको तो सर्वथा त्याग ही देना चाहिये

Sanskrit Verse:
न कामाननुरुद्ध्येत दुःखं कामेषु वै रतिः प्राप्यार्थमुपयुञ्जीत धर्मं कामान् विसर्जयेत् ॥


Example 2:
Meaning:
कारण कि यज्ञसे शेष बचे हुए अन्नको खानेवाले श्रेष्ठ पुरुष सब पापोंसे छूटते हैं और जो पापीलोग अपने शरीरपोषणके लिये ही पकाते हैं, वे तो पापको ही खाते हैं ॥

Sanskrit Verse:
यज्ञशिष्टाशिनः सन्तो मुच्यन्ते सर्वकिल्बिषैः भुञ्जते ते त्वयं पापा ये पचन्त्यात्मकारणात् ॥


Example 3:
Meaning:
वहाँसे सेनाविन्दुकी राजधानी देवप्रश्वमें आकर चतु रंगिणी सेनाके साथ शक्तिशाली अर्जुनने यहीं पड़ाव डाला ॥

Sanskrit Verse:
स देवप्रस्थमासाद्य सेनाविन्दोः पुरं प्रति बलेन चतुरङ्गेण निवेशमकरोत् प्रभुः ॥


--------------------------------------------------------------------------------------

split
train
--------------------------------------------------------------------------------------

text
<|im_start|>system
Objective:

Generate a sanskrit verse in Anuṣṭubh meter based on a given meaning or theme while strictly following the metrical (?and rhyme) rules.

Rules of Chandas:
Syllable Rules:
LAGHU vowels: अ, इ, उ, ऋ, ऌ
GURU vowels: आ, ई, ऊ, ॠ, ॡ, ए, ऐ, ओ, औ
HRASVA vowels: अ, इ, उ, ऋ, ऌ
DEERGHA vowels: आ, ई, ऊ, ॠ, ॡ, ए, ऐ, ओ, औ
Syllable is marked laghu, guru and hrasva, deergha based on the vowel it contains.
Syllable containing anusvāra("ं") or visarga("ः") is always marked as guru.
Syllable that is followed by a conjunct consonant (saṁyuktākṣara)
is always marked guru.

The goal is to generate Sanskrit verse that follows the anushtup meter rules for the given input translation.
RULES:
Verse Rules:
The verse contains 32 syllables/akshara and 4 padas in total.
The verse is divided in 2 lines, each containing 16 syllables.
Each line is divided into 2 padas (quartets),each containing exactly 8 syllables.
The fifth syllable of every pada must be LAGHU or short.
The sixth syllable of every pada must be GURU or long.
The seventh syllable of the second and the fourth pada must be HRASVA
The seventh syllable of the first and third paada must be DEERGHA

Language:
The generated verse must be in sanskrit in Devanagri script.
The Input given is in the hindi.
Do NOT include English words, transliteration, labels, or explanations.

Output Format:
(Output exactly 4 lines, one pada per line. Nothing else.)

Example 1:
Meaning:
इसलिये कामनाओं या भोगोंकी वृद्धिके लिये आग्रह नहीं रखना चाहिये भोगों में जो आसक्ति होती है, वह दुःखरूप ही है धन पाकर भी उसे धर्ममें ही लगा देना चाहिये काम भोगोंको तो सर्वथा त्याग ही देना चाहिये

Sanskrit Verse:
न कामाननुरुद्ध्येत दुःखं कामेषु वै रतिः प्राप्यार्थमुपयुञ्जीत धर्मं कामान् विसर्जयेत् ॥


Example 2:
Meaning:
कारण कि यज्ञसे शेष बचे हुए अन्नको खानेवाले श्रेष्ठ पुरुष सब पापोंसे छूटते हैं और जो पापीलोग अपने शरीरपोषणके लिये ही पकाते हैं, वे तो पापको ही खाते हैं ॥

Sanskrit Verse:
यज्ञशिष्टाशिनः सन्तो मुच्यन्ते सर्वकिल्बिषैः भुञ्जते ते त्वयं पापा ये पचन्त्यात्मकारणात् ॥


Example 3:
Meaning:
वहाँसे सेनाविन्दुकी राजधानी देवप्रश्वमें आकर चतु रंगिणी सेनाके साथ शक्तिशाली अर्जुनने यहीं पड़ाव डाला ॥

Sanskrit Verse:
स देवप्रस्थमासाद्य सेनाविन्दोः पुरं प्रति बलेन चतुरङ्गेण निवेशमकरोत् प्रभुः ॥<|im_end|>
<|im_start|>user
Meaning:
दिन और रात्रि उनके नेत्र हैं, वे दिव्यस्वरूप हैं वेद उनका अड्ग और श्रुतियाँ आभूषण हैं हविष्य नासिका, खुवा थूथन और सामवेदका गम्भीर घोष ही उनका स्वर है वे महान्‌ हैं ॥

Generate the Sanskrit verse.<|im_end|>
<|im_start|>assistant
अहोरात्रेक्षणो दिव्यो वेदाड़ः श्रुतिभूषण: आज्यनास: स्त्रुवातुण्डः सामघोषस्वनो महान्‌॥<|im_end|>

--------------------------------------------------------------------------------------

--------------------------------------------------------------------------------------
warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.
`logging_dir` is deprecated and will be removed in v5.2. Please set `TENSORBOARD_LOGGING_DIR` instead.
[unsloth.trainer|WARNING]Unsloth: packing=True ignored (hybrid linear-attention model).

Unsloth: Tokenizing ["text"]:   0%|          | 0/23316 [00:00<?, ? examples/s]
Unsloth: Tokenizing ["text"]:   4%|▍         | 1000/23316 [00:02<00:45, 487.69 examples/s]
Unsloth: Tokenizing ["text"]:   9%|▊         | 2000/23316 [00:04<00:43, 485.46 examples/s]
Unsloth: Tokenizing ["text"]:  13%|█▎        | 3000/23316 [00:06<00:41, 486.37 examples/s]
Unsloth: Tokenizing ["text"]:  17%|█▋        | 4000/23316 [00:08<00:39, 485.84 examples/s]
Unsloth: Tokenizing ["text"]:  21%|██▏       | 5000/23316 [00:10<00:38, 481.85 examples/s]
Unsloth: Tokenizing ["text"]:  26%|██▌       | 6000/23316 [00:12<00:35, 481.45 examples/s]
Unsloth: Tokenizing ["text"]:  30%|███       | 7000/23316 [00:14<00:33, 482.90 examples/s]
Unsloth: Tokenizing ["text"]:  34%|███▍      | 8000/23316 [00:16<00:31, 483.30 examples/s]
Unsloth: Tokenizing ["text"]:  39%|███▊      | 9000/23316 [00:18<00:29, 482.40 examples/s]
Unsloth: Tokenizing ["text"]:  43%|████▎     | 10000/23316 [00:20<00:27, 481.82 examples/s]
Unsloth: Tokenizing ["text"]:  47%|████▋     | 11000/23316 [00:22<00:25, 481.11 examples/s]
Unsloth: Tokenizing ["text"]:  51%|█████▏    | 12000/23316 [00:24<00:23, 481.09 examples/s]
Unsloth: Tokenizing ["text"]:  56%|█████▌    | 13000/23316 [00:26<00:21, 480.59 examples/s]
Unsloth: Tokenizing ["text"]:  60%|██████    | 14000/23316 [00:29<00:19, 479.96 examples/s]
Unsloth: Tokenizing ["text"]:  64%|██████▍   | 15000/23316 [00:31<00:17, 481.41 examples/s]
Unsloth: Tokenizing ["text"]:  69%|██████▊   | 16000/23316 [00:33<00:15, 481.99 examples/s]
Unsloth: Tokenizing ["text"]:  73%|███████▎  | 17000/23316 [00:35<00:13, 482.97 examples/s]
Unsloth: Tokenizing ["text"]:  77%|███████▋  | 18000/23316 [00:37<00:11, 483.17 examples/s]
Unsloth: Tokenizing ["text"]:  81%|████████▏ | 19000/23316 [00:39<00:08, 483.60 examples/s]
Unsloth: Tokenizing ["text"]:  86%|████████▌ | 20000/23316 [00:41<00:06, 482.90 examples/s]
Unsloth: Tokenizing ["text"]:  90%|█████████ | 21000/23316 [00:43<00:04, 482.51 examples/s]
Unsloth: Tokenizing ["text"]:  94%|█████████▍| 22000/23316 [00:45<00:02, 483.83 examples/s]
Unsloth: Tokenizing ["text"]:  99%|█████████▊| 23000/23316 [00:47<00:00, 483.68 examples/s]
Unsloth: Tokenizing ["text"]: 100%|██████████| 23316/23316 [00:48<00:00, 477.77 examples/s]
Unsloth: Tokenizing ["text"]: 100%|██████████| 23316/23316 [00:48<00:00, 479.48 examples/s]

Unsloth: Tokenizing ["text"]:   0%|          | 0/2918 [00:00<?, ? examples/s]
Unsloth: Tokenizing ["text"]:  34%|███▍      | 1000/2918 [00:02<00:03, 481.86 examples/s]
Unsloth: Tokenizing ["text"]:  69%|██████▊   | 2000/2918 [00:04<00:01, 483.80 examples/s]
Unsloth: Tokenizing ["text"]: 100%|██████████| 2918/2918 [00:06<00:00, 481.53 examples/s]
Unsloth: Tokenizing ["text"]: 100%|██████████| 2918/2918 [00:06<00:00, 473.50 examples/s]
[accelerate.utils.other|WARNING]Detected kernel version 5.4.0, which is below the recommended minimum of 5.5.0; this can cause the process to hang. It is recommended to upgrade the kernel to the minimum version or higher.

Map (num_proc=8):   0%|          | 0/23316 [00:00<?, ? examples/s]
Map (num_proc=8):   4%|▍         | 1000/23316 [00:00<00:15, 1400.50 examples/s]
Map (num_proc=8):  39%|███▊      | 9000/23316 [00:01<00:01, 9777.80 examples/s]
Map (num_proc=8):  73%|███████▎  | 16915/23316 [00:01<00:00, 14206.22 examples/s]
Map (num_proc=8): 100%|██████████| 23316/23316 [00:03<00:00, 6966.54 examples/s] 
Map (num_proc=8): 100%|██████████| 23316/23316 [00:03<00:00, 6855.97 examples/s]

Filter (num_proc=8):   0%|          | 0/23316 [00:00<?, ? examples/s]
Filter (num_proc=8):   4%|▍         | 1000/23316 [00:00<00:10, 2094.87 examples/s]
Filter (num_proc=8):  39%|███▊      | 9000/23316 [00:00<00:00, 16861.00 examples/s]
Filter (num_proc=8):  73%|███████▎  | 16915/23316 [00:00<00:00, 27404.57 examples/s]
Filter (num_proc=8): 100%|██████████| 23316/23316 [00:00<00:00, 23770.86 examples/s]
Unsloth: Removed 194 out of 23316 samples from train_dataset where all labels were -100 (no response marker found, usually truncation). This prevents NaN loss during training.

Map (num_proc=8):   0%|          | 0/2918 [00:00<?, ? examples/s]
Map (num_proc=8):  13%|█▎        | 365/2918 [00:00<00:02, 989.34 examples/s]
Map (num_proc=8): 100%|██████████| 2918/2918 [00:00<00:00, 7560.71 examples/s]
Map (num_proc=8): 100%|██████████| 2918/2918 [00:00<00:00, 4084.03 examples/s]

Filter (num_proc=8):   0%|          | 0/2918 [00:00<?, ? examples/s]
Filter (num_proc=8):  13%|█▎        | 365/2918 [00:00<00:02, 1135.67 examples/s]
Filter (num_proc=8): 100%|██████████| 2918/2918 [00:00<00:00, 5839.31 examples/s]
Unsloth: Removed 26 out of 2918 samples from eval_dataset where all labels were -100 (no response marker found, usually truncation). This prevents NaN loss during training.
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 248046, 'bos_token_id': None}.
==((====))==  Unsloth - 2x faster free finetuning | Num GPUs used = 1
   \\   /|    Num examples = 23,122 | Num Epochs = 5 | Total steps = 3,615
O^O/ \_/ \    Batch size per device = 16 | Gradient accumulation steps = 2
\        /    Data Parallel GPUs = 1 | Total batch size (16 x 2 x 1) = 32
 "-____-"     Trainable parameters = 159,383,552 of 27,055,382,016 (0.59% trained)

  0%|          | 0/3615 [00:00<?, ?it/s]Unsloth: Will smartly offload gradients to save VRAM!
Unsloth: Double buffering enabled (parallel H2D + compute) for backward pass.

  0%|          | 1/3615 [03:57<238:29:02, 237.56s/it]
                                                     
{'loss': '3.818', 'grad_norm': '3.531', 'learning_rate': '0', 'epoch': '0.001383'}

  0%|          | 1/3615 [03:57<238:29:02, 237.56s/it]
  0%|          | 2/3615 [05:42<160:11:25, 159.61s/it]
  0%|          | 3/3615 [07:19<131:20:45, 130.91s/it]
  0%|          | 4/3615 [08:43<112:43:37, 112.38s/it]
  0%|          | 5/3615 [10:34<112:08:28, 111.83s/it]
  0%|          | 6/3615 [12:14<108:20:16, 108.07s/it]
  0%|          | 7/3615 [13:50<104:11:09, 103.95s/it]
  0%|          | 8/3615 [15:32<103:38:45, 103.44s/it]
  0%|          | 9/3615 [16:57<97:39:37, 97.50s/it]  
  0%|          | 10/3615 [18:30<96:15:17, 96.12s/it]
                                                    
{'loss': '3.622', 'grad_norm': '2.747', 'learning_rate': '1.193e-05', 'epoch': '0.01383'}

  0%|          | 10/3615 [18:30<96:15:17, 96.12s/it]
  0%|          | 11/3615 [20:16<99:23:13, 99.28s/it]
  0%|          | 12/3615 [21:46<96:33:12, 96.47s/it]
  0%|          | 13/3615 [23:31<99:01:56, 98.98s/it]
  0%|          | 14/3615 [24:56<94:44:29, 94.72s/it]
  0%|          | 15/3615 [26:38<97:06:04, 97.10s/it]
  0%|          | 16/3615 [28:19<98:06:46, 98.14s/it]