PrepareBuddy IELTS-9B (Qwen3.5) — the most factually accurate

The largest model in the family. A fine-tune of Qwen3.5-9B (Apache-2.0) on PrepareBuddy's curated IELTS content. It generates IELTS practice across all four sections.

A content generator, not an assessment tool. It writes passages, transcripts, tasks, questions and answer keys. It does not score student work. A fine-tune of Qwen3.5-9B — not a from-scratch foundation model.

Where the 9B fits (the honest study result)

Part of our 2B/4B/9B study. Two things the size buys — and one it doesn't:

• ✅ Best world-knowledge / facts: from-scratch passages get names, dates and places right more often than the smaller models (e.g. correct "Central Asia" where a 2B wrote the wrong region).
• ✅ 100% completion answers verbatim-in-passage.
• ❗ Size did not buy better verdict reasoning — fine-tuning was flat (base 79% → ft 77%), and a fine-tuned 2B matched it on verdicts. So pick the 9B for fact-heavy from-scratch generation, not because it reasons better about verdicts.

Full numbers: technical report.

Where it fits best (real-world use cases)

Reach for the 9B when factual fidelity is the priority and you have a GPU:

• Fact-heavy from-scratch passages — science, history, geography: it gets names/dates/places right most often.
• A content factory on a GPU server — bulk-generate a question bank where passage accuracy matters most.
• The top-quality tier behind a paid product, paired with grounding + verification.
• Sentence/summary completion — 100% in-passage answers.

Not the best pick for laptops/edge or cost-sensitive serving (→ 2B/4B), or if you expect better verdict reasoning — size didn't buy that (a fine-tuned 2B matched it).

Pros & cons

✅ Pros	⚠️ Cons
Best facts in the family; 100% completion	Heaviest — ~18 GB VRAM bf16 (quantise to ~5–6 GB)
Strong Writing / Speaking / Listening	MCQ answer-position skews to "B" — fix at serving
Most reliable from-scratch passages	Verdict reasoning no better than the 2B (~8% easy / ≈25% varied)
Apache-2.0; the quality ceiling of the set	Best on a GPU/server, not a laptop

Links

• 🧠 Models: ielts-2b · ielts-4b · ielts-9b
• 💻 Apple Silicon / LM Studio (MLX): ielts-2b-mlx · ielts-4b-mlx
• 🌐 Try the live demo: Hugging Face Space
• 📄 Full technical report & findings: ielts-qwen3.5

What it generates

Section	Types	Output
Reading	TFNG, YNNG, MCQ, Sentence/Summary Completion, Matching, Long-form	passage + questions + answer key with justifications
Writing	Task 1 (chart), Task 2 (essay)	task prompt + word limit + timing (+ chart data for T1)
Listening	dialogue/monologue	transcript + questions + answer key (text for downstream TTS)
Speaking	Part 1, 2, 3	examiner question / cue card + model answer

Prompt format (not a chat model — use the tag prefix)

<TEST=IELTS><SECTION=READING><TYPE=TFNG><DIFF=medium><TOPIC=ocean currents> Generate a short passage with 4 True/False/Not Given statements and an answer key.

• SECTION = READING | WRITING | LISTENING | SPEAKING
• TYPE (Reading) = TFNG | YNNG | MCQ | MCQ_MULTI | SENTENCE_COMPLETION | SUMMARY_COMPLETION | MATCHING_HEADINGS | MATCHING_FEATURES | MATCHING_ENDINGS | LONGFORM; (Writing) TASK1 | TASK2; (Speaking) PART1 | PART2 | PART3; (Listening) LISTENING
• DIFF = easy | medium | hard
• Serve with enable_thinking=False — for this task, reasoning mode lowers verdict accuracy.

Examples (the family's shared output format; answer keys verified)

These illustrate the format all three models produce — the 9B generates it with the highest factual accuracy and 100% completion grounding. Reading examples use the recommended grounded approach (generate against a real passage), so every answer key here was checked against its passage. (The 9B needs a GPU to run — ~18 GB in bf16 — so we show the shared verified format rather than a laptop generation.)

Writing — Task 2 (temp 0.7)

Input

<TEST=IELTS><SECTION=WRITING><TYPE=TASK2><DIFF=medium><TOPIC=whether university education should be free> Generate an IELTS Academic Writing Task 2 question with word limit and timing.

Output

TASK:
Some people believe that university education should be free for all students who wish to
pursue a degree, while others think this would result in very high costs.
To what extent do you agree or disagree?

WORD LIMIT: 250 words   TIME: 40 minutes

Reading — True/False/Not Given (grounded, temp 0.3 — keys verified)

Input

<TEST=IELTS><SECTION=READING><TYPE=TFNG><DIFF=medium><TOPIC=coral reefs> Using ONLY the passage below, write True/False/Not Given statements with an answer key. Do not write a new passage.
PASSAGE: Coral reefs … are built by tiny animals called polyps, which secrete calcium carbonate to form hard skeletons. Reefs grow only in warm, shallow, sunlit water … The Great Barrier Reef, off Australia, is the largest reef system in the world.

Output

QUESTIONS:
1. Coral reefs are built by tiny animals called polyps that secrete calcium carbonate.
2. Coral reefs grow in cold, deep water because the polyps depend on algae for food.
3. The Great Barrier Reef is the largest reef system on Earth.

ANSWER KEY:
1. TRUE  — "built by tiny animals called polyps, which secrete calcium carbonate".
2. FALSE — the passage says reefs "grow only in warm, shallow, sunlit water", not cold deep water.
3. TRUE  — "The Great Barrier Reef … is the largest reef system in the world".

Reading — Multiple Choice (grounded, temp 0.3 — key verified)

Output

QUESTION — Choose the correct letter, A, B, C or D.
Why do coral polyps depend on algae living in their tissues?
 A. The algae provide calcium carbonate to build the skeleton.
 B. Algae help polyps survive in cold, deep water.
 C. The algae supply food to the polyps through photosynthesis.
 D. Algae protect polyps from bleaching.

ANSWER KEY:
1. C — the passage states polyps "depend on algae living in their tissues for food"; the other options are not stated.

(⚠️ MCQ correct-answer letter skews toward "B" on the 9B — spread answer positions at serving, or only publish pre-checked MCQ.)

Reading — Sentence Completion (grounded, temp 0.3 — 9B answers 100% in-passage)

Output

QUESTIONS — Complete each sentence with ONE word from the passage.
1. Reefs grow only in warm, shallow water because the polyps depend on algae for _____.   → food
2. When water becomes too warm, the polyps expel the algae and turn white — a process known as _____.   → bleaching

Supported types per section — and how to prompt each

Section · Type	Prompt <TYPE=…>	Temp	What you get
Reading · True/False/Not Given	TFNG	0.3	passage + statements + key
Reading · Yes/No/Not Given	YNNG	0.3	opinion passage + statements + key
Reading · Multiple choice	MCQ / MCQ_MULTI	0.3	passage + A–D question(s) + key (watch the "B" skew)
Reading · Sentence/Summary completion	SENTENCE_COMPLETION / SUMMARY_COMPLETION	0.3	gap items + key (9B: 100% in-passage)
Reading · Matching	MATCHING_*	0.5	matching task + key (experimental)
Reading · Long-form	LONGFORM	0.6	~600-word passage + mixed questions + key
Writing · Task 1 / Task 2	TASK1 / TASK2	0.7	task + word limit + timing
Speaking · Part 1/2/3	PART1 / PART2 / PART3	0.7	examiner question / cue card + model answer
Listening	LISTENING	0.7	transcript + questions + key

Tip — for dependable Reading keys, generate grounded: prepend a real passage and add "Using ONLY the passage below … Do not write a new passage."

Generating a full exam section (one passage → all question types)

Generate one passage, then each question type against it:

1. <…TYPE=LONGFORM…> Write ONLY a ~600-word IELTS reading passage. No questions.
2. For each type: Using ONLY the passage below, write 5 TFNG statements with an answer key. Do not write a new passage.\nPASSAGE:\n<passage>
3. Concatenate → a real-exam-style section. The 9B's strong facts make its from-scratch passages the most reliable base. (The demo Space does this.)

Usage (transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "preparebuddy/ielts-9b"
tok = AutoTokenizer.from_pretrained(repo)
# ~18 GB in bf16 (24 GB+ GPU). For smaller GPUs, load 4-bit (~5–6 GB):
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto").eval()
SYSTEM = ("You generate authentic IELTS Academic practice content across reading, writing, "
          "listening, and speaking. Produce passages, transcripts, tasks, questions, and answer "
          "keys or model answers as appropriate to the section. Use IELTS-style register: "
          "academic, neutral, factually plausible. This is content generation, not assessment.")
user = "<TEST=IELTS><SECTION=READING><TYPE=TFNG><DIFF=medium><TOPIC=solar power> Generate a short passage with 4 True/False/Not Given statements and an answer key."
inp = tok.apply_chat_template([{"role":"system","content":SYSTEM},{"role":"user","content":user}],
        add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inp, max_new_tokens=900, do_sample=True, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))

Settings: temp 0.3 for verdicts (TFNG/YNNG/MCQ), 0.7 for passages/writing/speaking; top_p 0.9; one SECTION+TYPE per call. VRAM: ~18 GB bf16; quantise (load_in_4bit=True) to ~5–6 GB for normal GPUs. The 9B's home is code/server or a GPU Space rather than a laptop.

Recommended architecture for reliable output (important)

The 9B is a strong drafter with the family's best facts. For dependable answer keys, run it as a system:

1. Ground — generate against a real passage (the 9B's strong facts already make from-scratch passages reliable, but grounding removes residual risk).
2. Verify — re-check each answer key with an independent judge (a trained 2B is a cheap, effective verifier) and flag disagreements.
3. Review/regenerate the small flagged minority.

Measured end-to-end: raw grounded generation ≈ 75% → ≈ 85–90% with this verify loop.

Strengths & honest limits (9B)

• ✅ Best facts in the family; 100% completion; strong Writing/Speaking/Listening; 0 non-English-token leak.
• ⚠️ MCQ answer-position skews to "B" (in our sample the correct answer was B 7/7) — spread the position at serving, or only publish pre-checked MCQ.
• ⚠️ Verdict reasoning is no better than the smaller models — ~8% logic slips on easy items (≈25% on varied from-scratch); use grounding + verification.
• ⚠️ Heaviest to run — ~18 GB VRAM (bf16); best on a GPU/server, not a laptop.
• Listening/Speaking output is text (for downstream TTS); no audio. Not an assessment tool.

Training

LoRA fine-tune of Qwen3.5-9B (bf16; r16/α32; completion-only loss; enable_thinking=False; 2 epochs, lr 1e-4), trained on an NVIDIA L40S (48 GB) cloud GPU, on 1,438 curated + balanced examples (≈⅓ NOT GIVEN in verdict types). Dataset not released (proprietary). Full method, hardware and results: technical report.

License

Apache-2.0, inheriting from Qwen3.5-9B. Free to use, modify, distribute (incl. commercially); retain attribution to the base model and PrepareBuddy.

---

The 2B / 4B / 9B family — pick the right one

	ielts-2b	ielts-4b ⭐	ielts-9b
Best for	cheapest; best verdict judge/verifier	balanced general use	best facts (from scratch)
Verdict accuracy (fine-tuned)¹	80%	74%	77%
Completion answers in-passage	⚠️ 37%	✅ 100%	✅ 100%
Facts in from-scratch passages	weakest	good	✅ best
MCQ answer-position	ok	ok	⚠️ skews "B"
Size (bf16)	~5 GB	~9 GB	~18 GB
Use with grounding	strongly	recommended	recommended

¹ greedy, 101-item held-out gold. Fine-tuning's benefit is inversely proportional to base capability — it transformed the 2B (+40) and was flat on the 4B/9B. Full method + findings + tables: technical report.

Getting better results: grounding + the re-checking loop (the biggest quality lever)

1. Ground — generate against a real passage so facts come from the source:

Using ONLY the passage below, write 4 True/False/Not Given statements with an answer key. Do NOT write a new passage.
PASSAGE: <your real passage>

2. Re-check (verify) — independently re-judge each answer key, flag disagreements:

for statement in generated_statements:
    verdict = judge(model, passage, statement)      # TRUE / FALSE / NOT GIVEN
    if verdict != generated_key[statement]:
        flag_for_review_or_regenerate(statement)    # the verifier catches ~75-80% of errors

A trained 2B catches ~75–80% of verdict errors as a verifier (cheap); any 4B+ works too. 3. Review / regenerate the flagged minority. Measured: ≈ 75% raw → ≈ 85–90% with this loop.

Prompt tips

• Always use the tag prefix <TEST=IELTS><SECTION=…><TYPE=…><DIFF=…><TOPIC=…> — it's not a chat model.
• Temperature: 0.3 for verdicts, 0.7 for passages/writing/speaking; top_p 0.9.
• enable_thinking=False — reasoning mode lowers verdict accuracy.
• One SECTION+TYPE per call; build a full section by generating each type against one shared passage.