Update README.md
Browse files
README.md
CHANGED
|
@@ -12,16 +12,24 @@ tags:
|
|
| 12 |
|
| 13 |
# SAT Tutor — Qwen2.5-7B LoRA Adapter
|
| 14 |
|
|
|
|
|
|
|
| 15 |
A fine-tuned SAT tutor built on **Qwen/Qwen2.5-7B-Instruct** with QLoRA (4-bit). It answers SAT Math and Reading & Writing questions with step-by-step explanations, and can also give hint-only nudges (first step only) so the student finishes the problem themselves.
|
| 16 |
|
| 17 |
## Training
|
| 18 |
|
| 19 |
- **Method:** QLoRA — 4-bit NF4, LoRA rank 16 / alpha 32 (~40.4M trainable params, 0.53% of 7.6B)
|
|
|
|
| 20 |
- **Data:** 15,000 examples from public SAT sources:
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
- **Format:** ~75% full step-by-step solutions ending in `Answer: X`; ~25% hint-only responses
|
|
|
|
| 25 |
- **Run:** 1 epoch, 938 steps, 1024-token context, fp16, 2× Kaggle T4 (~7h)
|
| 26 |
|
| 27 |
## Evaluation (held-out official questions)
|
|
@@ -50,10 +58,15 @@ model = PeftModel.from_pretrained(model, "agk4444/sat-tutor-qwen2.5-7b")
|
|
| 50 |
model.eval()
|
| 51 |
|
| 52 |
msgs = [{"role": "user", "content":
|
| 53 |
-
|
| 54 |
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
|
| 55 |
inp = tok(prompt, return_tensors="pt").to(model.device)
|
| 56 |
out = model.generate(inp["input_ids"], attention_mask=inp.get("attention_mask"),
|
| 57 |
max_new_tokens=400, do_sample=False,
|
| 58 |
pad_token_id=tok.eos_token_id)
|
| 59 |
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
# SAT Tutor — Qwen2.5-7B LoRA Adapter
|
| 14 |
|
| 15 |
+
Developed by **AGK FIRE INC**.
|
| 16 |
+
|
| 17 |
A fine-tuned SAT tutor built on **Qwen/Qwen2.5-7B-Instruct** with QLoRA (4-bit). It answers SAT Math and Reading & Writing questions with step-by-step explanations, and can also give hint-only nudges (first step only) so the student finishes the problem themselves.
|
| 18 |
|
| 19 |
## Training
|
| 20 |
|
| 21 |
- **Method:** QLoRA — 4-bit NF4, LoRA rank 16 / alpha 32 (~40.4M trainable params, 0.53% of 7.6B)
|
| 22 |
+
|
| 23 |
- **Data:** 15,000 examples from public SAT sources:
|
| 24 |
+
|
| 25 |
+
- `ndavidson/sat-math-chain-of-thought` — step-by-step math solutions (deduplicated, correct-only)
|
| 26 |
+
|
| 27 |
+
- `betterMateusz/SAT_Writting_Reading_Assessment_Question_Bank` — official-style reading/writing questions with rationales
|
| 28 |
+
|
| 29 |
+
- `emozilla/sat-reading` — passage comprehension
|
| 30 |
+
|
| 31 |
- **Format:** ~75% full step-by-step solutions ending in `Answer: X`; ~25% hint-only responses
|
| 32 |
+
|
| 33 |
- **Run:** 1 epoch, 938 steps, 1024-token context, fp16, 2× Kaggle T4 (~7h)
|
| 34 |
|
| 35 |
## Evaluation (held-out official questions)
|
|
|
|
| 58 |
model.eval()
|
| 59 |
|
| 60 |
msgs = [{"role": "user", "content":
|
| 61 |
+
"SAT Math practice question:\n\nIf 3x + 5 = 20, what is x?\nA) 3 B) 5 C) 7 D) 15"}]
|
| 62 |
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
|
| 63 |
inp = tok(prompt, return_tensors="pt").to(model.device)
|
| 64 |
out = model.generate(inp["input_ids"], attention_mask=inp.get("attention_mask"),
|
| 65 |
max_new_tokens=400, do_sample=False,
|
| 66 |
pad_token_id=tok.eos_token_id)
|
| 67 |
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))
|
| 68 |
+
```
|
| 69 |
+
|
| 70 |
+
---
|
| 71 |
+
|
| 72 |
+
© 2026 AGK FIRE INC. Released under Apache 2.0.
|