Hugging Face's logo

kartikey31
/
txn-parser

Text Generation
PEFT
Safetensors
GGUF
English
Hindi
lora
qlora
transaction-parser
on-device
conversational
Model card Files
xet
 Community
kartikey31 commited on
Commit
9a42d8a
·
verified ·
1 Parent(s): 362235a

Upload gemma-3-270m/README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. gemma-3-270m/README.md +71 -5
gemma-3-270m/README.md CHANGED
@@ -55,8 +55,37 @@ sibling fine-tunes of other base models trained on the same data.
55
  | Eval batch size | 4 |
56
  | Max seq length | 1024 |
57
  | Learning rate | 2e-4 (warmup 3%) |
58
- | Started | 2026-05-22T19:10:02.542366+00:00 |
59
- | Finished | 2026-05-22T20:40:44.106745+00:00 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
60
 
61
  ## Download a single GGUF
62
 
@@ -66,12 +95,49 @@ huggingface-cli download kartikey31/txn-parser \
66
  --local-dir .
67
  ```
68
 
69
- ## Inference (GGUF, llama.cpp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
70
 
71
  ```bash
72
  ./llama-cli -m txn-parser-gemma-3-270m-Q4_K_M.gguf \
73
  --grammar-file scripts/grammar.gbnf \
74
- -p "Spent 200 on chai" -n 256
 
75
  ```
76
 
77
  ## Reproduce
@@ -83,4 +149,4 @@ python scripts/train_and_publish.py --only gemma-3-270m
83
  ```
84
 
85
  ---
86
- *Auto-published by `scripts/train_and_publish.py` on 2026-05-22T20:40:44.198456+00:00.*
 
55
  | Eval batch size | 4 |
56
  | Max seq length | 1024 |
57
  | Learning rate | 2e-4 (warmup 3%) |
58
+ | Started | |
59
+ | Finished | |
60
+
61
+ ## System prompt (use this EXACTLY)
62
+
63
+ The model was trained with one specific system prompt and Gemma/Smol/Qwen
64
+ chat template. If you paraphrase the prompt or skip the chat template,
65
+ quality degrades quickly. Copy-paste this verbatim into your inference
66
+ client (no leading/trailing whitespace, no edits):
67
+
68
+ ```text
69
+ You convert voice-transcribed transaction descriptions into structured JSON.
70
+
71
+ Output ONLY a JSON object with this schema, no other text:
72
+ {"transactions":[{"amount":<number>,"currency":"INR"|"USD","item":"<lowercase singular noun phrase>","category":"<enum>","type":"expense"|"income"}]}
73
+
74
+ Categories: Food, Drinks, Groceries, Transport, Shopping, Entertainment, Bills, Health, Education, Personal, Gifts, Income, Other.
75
+
76
+ Rules:
77
+ - Currency defaults to INR. Use USD only when the input explicitly says "dollars" or contains "$".
78
+ - Amounts: "k" = ×1000, "hazaar" = ×1000, "sau" = ×100, "lakh" = ×100000. Convert number-words ("five hundred") to digits.
79
+ - type is "expense" by default; "income" only for explicit salary, cashback, refund, gift received, payment received.
80
+ - For disfluencies and corrections ("500 wait no 600"), output the CORRECTED amount only.
81
+ - For ambiguous items ("that thing", "stuff"), use item "unspecified" and category "Other".
82
+ - Item field: lowercase singular noun phrase ("uber ride", "beer", "chai" — not "Beers" or "Uber").
83
+ - Multi-transaction inputs become multiple array entries in spoken order.
84
+ - Category heuristics: uber/ola/auto/petrol/bus/metro → Transport; beer/wine/chai/coffee/juice → Drinks; rent/electricity/wifi/recharge/gas → Bills; movie/netflix/concert → Entertainment; doctor/medicine/hospital → Health.
85
+ ```
86
+
87
+ Source of truth: [`scripts/_lib.py`](https://github.com/kartikeychoudhary/txn-parser/blob/main/scripts/_lib.py)
88
+ constant `SYSTEM_PROMPT`. Don't retype it — pull from `_lib.py` or this README.
89
 
90
  ## Download a single GGUF
91
 
 
95
  --local-dir .
96
  ```
97
 
98
+ ## Inference (Python, llama-cpp-python)
99
+
100
+ ```python
101
+ from llama_cpp import Llama
102
+
103
+ SYSTEM_PROMPT = '''You convert voice-transcribed transaction descriptions into structured JSON.
104
+
105
+ Output ONLY a JSON object with this schema, no other text:
106
+ {"transactions":[{"amount":<number>,"currency":"INR"|"USD","item":"<lowercase singular noun phrase>","category":"<enum>","type":"expense"|"income"}]}
107
+
108
+ Categories: Food, Drinks, Groceries, Transport, Shopping, Entertainment, Bills, Health, Education, Personal, Gifts, Income, Other.
109
+
110
+ Rules:
111
+ - Currency defaults to INR. Use USD only when the input explicitly says "dollars" or contains "$".
112
+ - Amounts: "k" = ×1000, "hazaar" = ×1000, "sau" = ×100, "lakh" = ×100000. Convert number-words ("five hundred") to digits.
113
+ - type is "expense" by default; "income" only for explicit salary, cashback, refund, gift received, payment received.
114
+ - For disfluencies and corrections ("500 wait no 600"), output the CORRECTED amount only.
115
+ - For ambiguous items ("that thing", "stuff"), use item "unspecified" and category "Other".
116
+ - Item field: lowercase singular noun phrase ("uber ride", "beer", "chai" — not "Beers" or "Uber").
117
+ - Multi-transaction inputs become multiple array entries in spoken order.
118
+ - Category heuristics: uber/ola/auto/petrol/bus/metro → Transport; beer/wine/chai/coffee/juice → Drinks; rent/electricity/wifi/recharge/gas → Bills; movie/netflix/concert → Entertainment; doctor/medicine/hospital → Health.'''
119
+
120
+ llm = Llama(
121
+ model_path="txn-parser-gemma-3-270m-Q4_K_M.gguf",
122
+ n_gpu_layers=-1, n_ctx=2048,
123
+ )
124
+ out = llm.create_chat_completion(
125
+ messages=[
126
+ {"role": "system", "content": SYSTEM_PROMPT},
127
+ {"role": "user", "content": "200 ka samosa"},
128
+ ],
129
+ temperature=0.0,
130
+ )
131
+ print(out["choices"][0]["message"]["content"])
132
+ ```
133
+
134
+ ## Inference (CLI, llama.cpp)
135
 
136
  ```bash
137
  ./llama-cli -m txn-parser-gemma-3-270m-Q4_K_M.gguf \
138
  --grammar-file scripts/grammar.gbnf \
139
+ --system-prompt "$(cat system_prompt.txt)" \
140
+ -p "200 ka samosa" -n 256
141
  ```
142
 
143
  ## Reproduce
 
149
  ```
150
 
151
  ---
152
+ *Auto-published by `scripts/train_and_publish.py` on 2026-05-22T21:33:31.274220+00:00.*