A lightweight, no-strings-attached Chain-of-Thought framework for your LLM, ensuring reliable results for bulk input requests stored in CSV
/ JSONL
/ sqlite
.
It allows applying series of prompts formed into schema
(See related section)
- ✅ No-strings: you're free to LLM dependencies and flexible
venv
customization. - ✅ Provides iterator over infinite amount of input contexts served in
CSV
/JSONL
. - ✅ Progress caching: withstanding exception during LLM calls by using
sqlite3
engine for caching LLM answers; - ✅ Support schemas descriptions for Chain-of-Thought concept.
pip install bulk-chain
To declare Chain-of-Though (CoT) schema, this project exploits JSON
format.
This format adopts name
field for declaring a name and schema
is a list of CoT instructions for the Large Language Model.
Each step represents a dictionary with prompt
and out
keys that corresponds to the input prompt and output variable name respectively.
All the variable names are expected to be mentioned in {}
.
Below, is an example on how to declare your own schema:
{
"name": "schema-name",
"schema": [
{"prompt": "Given the question '{text}', let's think step-by-step.",
"out": "steps"},
{"prompt": "For the question '{text}' the reasoining steps are '{steps}'. what would be an answer?",
"out": "answer"},
]
}
Another templates are available here.
Just three simple steps:
- Define your CoT Schema, or fetch it as shown below:
!wget https://raw.githubusercontent.com/nicolay-r/bulk-chain/refs/heads/master/ext/schema/default.json
- Fetch or write your own model or pick the one preset here:
!wget https://raw.githubusercontent.com/nicolay-r/bulk-chain/refs/heads/master/ext/flan_t5.py
- Launch inference in (chat mode):
!python -m bulk_chain.infer \
--schema "default.json" \
--adapter "dynamic:flan_t5.py:FlanT5" \
%% \
--device "cpu" \
--temp 0.1
All you have to do is to implement BaseLM
class, that includes:
__init__
-- for initialization;ask(prompt)
-- infer your model with the givenprompt
.
See examples with models here.