85 lines
2.4 KiB
Markdown
85 lines
2.4 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
base_model: Qwen/Qwen3-1.7B
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
datasets:
|
||
|
|
- cs-552-2026-aaty/sft_mixture
|
||
|
|
tags:
|
||
|
|
- qwen3
|
||
|
|
- general-knowledge
|
||
|
|
- reasoning
|
||
|
|
- sft
|
||
|
|
- cs-552
|
||
|
|
- mnlp
|
||
|
|
metrics:
|
||
|
|
- accuracy
|
||
|
|
---
|
||
|
|
|
||
|
|
# general_knowledge_model
|
||
|
|
|
||
|
|
General-knowledge specialty model for team **AATY**, CS-552 MNLP (EPFL).
|
||
|
|
It is `Qwen/Qwen3-1.7B` supervised fine-tuned on a general-knowledge mixture
|
||
|
|
to answer closed-book factual and reasoning questions across the sciences,
|
||
|
|
humanities, and geography.
|
||
|
|
|
||
|
|
## Model details
|
||
|
|
|
||
|
|
- **Base model:** `Qwen/Qwen3-1.7B`
|
||
|
|
- **Post-training:** supervised fine-tuning (LoRA adapter, merged back into the
|
||
|
|
base weights)
|
||
|
|
- **Domain:** general knowledge, multiple-choice, scored at pass@1
|
||
|
|
- **Format:** vLLM-loadable safetensors with `config.json`,
|
||
|
|
`generation_config.json`, and a tokenizer `chat_template`
|
||
|
|
|
||
|
|
## Output contract
|
||
|
|
|
||
|
|
The model writes its reasoning and then wraps the final answer in `\boxed{...}`.
|
||
|
|
For multiple-choice items the boxed content is the letter of the chosen option,
|
||
|
|
and option counts can range from 2 to 20.
|
||
|
|
|
||
|
|
```
|
||
|
|
Q: Which planet is closest to the Sun?
|
||
|
|
A) Venus
|
||
|
|
B) Mercury
|
||
|
|
C) Mars
|
||
|
|
D) Earth
|
||
|
|
A: ...reasoning... \boxed{B}
|
||
|
|
```
|
||
|
|
|
||
|
|
## Thinking mode
|
||
|
|
|
||
|
|
This model runs in **thinking mode**: it emits a `<think>...</think>` reasoning
|
||
|
|
block before the final `\boxed{...}` answer. Thinking is forced on inside the
|
||
|
|
chat template, because the evaluation passes only
|
||
|
|
`tokenizer.apply_chat_template(messages, add_generation_prompt=True)` with no
|
||
|
|
`enable_thinking` argument, so the template default is the only signal honored.
|
||
|
|
|
||
|
|
The relevant line in `chat_template.jinja`:
|
||
|
|
|
||
|
|
```jinja
|
||
|
|
{%- set enable_thinking = true %}
|
||
|
|
```
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
|
|
|
||
|
|
tok = AutoTokenizer.from_pretrained("cs-552-2026-aaty/general_knowledge_model")
|
||
|
|
model = AutoModelForCausalLM.from_pretrained("cs-552-2026-aaty/general_knowledge_model")
|
||
|
|
|
||
|
|
messages = [{"role": "user", "content": "What is the capital of Australia?"}]
|
||
|
|
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
|
|
inputs = tok(prompt, return_tensors="pt").to(model.device)
|
||
|
|
out = model.generate(**inputs, max_new_tokens=512)
|
||
|
|
print(tok.decode(out[0], skip_special_tokens=True))
|
||
|
|
```
|
||
|
|
|
||
|
|
## Training data
|
||
|
|
|
||
|
|
Supervised fine-tuning on `cs-552-2026-aaty/sft_mixture`, the chat-formatted
|
||
|
|
mixture built from public QA and knowledge datasets. See the team data pipeline
|
||
|
|
in `code/data/` for the exact sources and filters.
|