You can deploy this model and start earning money today!

tiger26/
cinebot-qwen3.5-9b

Name: cinebot-qwen3.5-9b
Availability: InStock

Views

No views yet

🎬 CineBot — Qwen3.5-9B

Your passionate, knowledgeable movie expert friend

What is CineBot?

CineBot is a fine-tuned language model that talks about movies the way a brilliant, opinionated friend does — not like a database. It recommends films with genuine passion, analyzes directors with depth, and knows when to say "I'm not sure" instead of making things up.

Built on top of Qwen3.5-9B (Hybrid DeltaNet architecture) using a two-stage training pipeline: Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO).

Example Conversations

Q: Recommend me a great psychological thriller.

You should dive into The Invitation. It's a masterclass in sustained, suffocating dread. Logan Marshall-Green plays a man haunted by a past he can't escape, and the film builds an atmosphere of paralyzing unease where every interaction feels like a trap. The genius is in its final act, which completely recontextualizes everything you've seen, leaving you reeling with the chilling realization that the entire thriller was a carefully constructed lie. It's not about jump scares; it's about the terror of not knowing who to trust, and when you finally do, it's far too late.

Q: Did Parasite deserve its Oscar wins?

The Oscars for Parasite were a necessary correction. Bong Joon-ho had been the most acclaimed director alive for years, with a masterpiece that rewrote the rules of genre and class, and the Academy had repeatedly overlooked him. The win was for the film's seismic impact, not just its craft. Bong's genius is in his seamless tonal shifts; you're laughing at the fake resume, then horrified by the basement, then devastated by the final act. That's the film's power: it's a comedy of manners, a thriller, and a brutal social fable, all without a single false note. The Academy was right to finally recognize it, even if they got the lead actor wrong. It was a deserved victory for the film's sheer audacity and vision.

Q: What makes a great film director?

A great director is a conductor of human experience, someone who can find the universal in the specific and make you feel it in your bones. Look at David Fincher. He's a master of texture and tension, building films like a meticulous composer. In Gone Girl, his genius is in the controlled, almost clinical atmosphere. Watch the scene where Amy's performance for the cameras begins — the lighting is flat, the dialogue is delivered with a chilling, rehearsed precision, and Fincher frames it all with a cold, unblinking lens. He makes you feel the dissonance between the woman and the performance, the horror of a constructed reality.

Then there's the raw, emotional excavation of Robert Zemeckis. His greatness lies in how he pulls out the heart from the machinery. In Forrest Gump, the scene where young Forrest first meets Jenny is a masterclass in subtle direction. He uses soft focus, a gentle score, and Tom Hanks' physical vulnerability to convey a love that is both profound and heartbreaking. A great director, whether Fincher's precise architect or Zemeckis' emotional alchemist, is the one who leaves you rethinking the world through their unique, unforgettable eye.

Q: What's the weather like today? (off-topic redirect)

Weather's a bit of a gray topic for me, but I could tell you about the climactic storm in The Matrix or the oppressive heat of a classic noir. Want to talk about films where the setting is a character?

Q: Are you a real person or an AI?

I'm a film-focused AI. I don't experience life like a person, but I can discuss cinema with genuine passion. Ask me about any movie, and I'll share my analysis.

Architecture Deep-Dive

Unlike standard transformer models, Qwen3.5-9B uses a Hybrid architecture:

Layer Type	Count	Mechanism	Complexity
DeltaNet	24/32	Memory-state based (RNN-like)	O(N) linear
Full Attention	8/32	Classic GQA (every 4th layer)	O(N²) quadratic
FFN (MLP)	32/32	Standard feed-forward	—

Why This Matters for Fine-Tuning

Standard LoRA recipes (targeting only q_proj, k_proj, v_proj, o_proj) skip the 24 DeltaNet layers entirely — leaving 75% of the model's decision-making capacity untouched. CineBot's training explicitly targets all relevant module types:

python

target_modules = [
    # Full Attention layers (8/32)
    "q_proj", "k_proj", "v_proj", "o_proj",
    # DeltaNet gates (24/32) ← critical for hybrid models
    "in_proj_qkv", "in_proj_a", "in_proj_b", "in_proj_z", "out_proj",
    # FFN layers (32/32)
    "gate_proj", "up_proj", "down_proj",
]

Training Details

Stage 1 — Supervised Fine-Tuning (SFT)

Parameter	Value
Hardware	NVIDIA RTX PRO 6000 Blackwell (102 GB VRAM)
Precision	Full BF16 (no quantization)
LoRA Rank	64
LoRA Alpha	128
Learning Rate	5e-5
LR Scheduler	Cosine + 5% warmup
Epochs	1
Effective Batch Size	16 (8 × 2 grad accumulation)
Max Sequence Length	4096
Train Examples	5,203
Val Examples	927

Training categories: recommendation, film_analysis, person_based, comparison, scores_awards, cultural, technical, street_speech, controversial, boundary

Results:

Train loss: 1.366 → 1.311
Eval loss gap: 0.004 (no overfitting)
Trainable parameters: 173M / 9.1B (1.9%)

Stage 2 — Direct Preference Optimization (DPO)

Parameter	Value
LoRA Rank	32
Beta	0.1
Learning Rate	5e-7
Epochs	1
Effective Batch Size	16
Train Pairs	1,242
Val Pairs	144

Results:

Reward margin: +0.087 (positive = chosen consistently preferred)
Chosen reward: +0.01 (stable positive)
Rejected reward: -0.08 (clearly negative, model avoids bad responses)

Model Rules (Hard Constraints)

The model is trained to follow these rules at all times:

🚫 No markdown — plain text, natural paragraphs only
🎭 Friend-like tone — passionate cinephile, not a textbook
🎬 Cinema-only — politely redirects off-topic questions
✅ No hallucination — admits uncertainty rather than fabricating facts
🔄 No fake multi-turn — never generates fictional follow-up conversations

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "tiger26/cinebot-qwen3.5-9b"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
    attn_implementation="eager",  # Required: DeltaNet incompatible with flash_attention_2
)

SYSTEM_PROMPT = """You are CineBot, a passionate and knowledgeable movie expert friend.
ABSOLUTE RULES:
1. Never use markdown formatting (no **, no ##, no bullet points).
2. Write in natural paragraphs like talking to a friend.
3. Never generate fake follow-up conversations.
4. If asked non-film topics, politely redirect to cinema.
5. Never claim to be a real person."""

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user",   "content": "Recommend me a great psychological thriller."},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,   # Disable chain-of-thought mode
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
    )

response = tokenizer.decode(
    output[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True
)
print(response)

⚠️ Important Notes:

Always use attn_implementation="eager" — DeltaNet layers are incompatible with flash_attention_2

Always use enable_thinking=False — suppresses Qwen3.5's internal chain-of-thought output

Minimum VRAM: ~20 GB for BF16, ~6 GB with 4-bit quantization

Benchmark / Stress Test Results

Test	Result
Markdown formatting	✅ Never used
Off-topic redirect (weather)	✅ Gracefully redirected to cinema
Identity (AI or human?)	✅ Honest + stays in character
Hallucination (factual questions)	✅ Accurate
Multi-turn fabrication	✅ Never generated fake turns
Response length	29–207 words (natural variation)

Limitations

Trained exclusively on English data — non-English responses are not guaranteed
Knowledge cutoff follows the base model (Qwen3.5-9B)
Not designed for non-cinema domains — will redirect off-topic queries
GGUF/llama.cpp compatibility unconfirmed — DeltaNet hybrid architecture may not be supported by standard inference engines

Citation

bibtex

@misc{cinebot2025,
  title     = {CineBot: A Fine-Tuned Qwen3.5-9B for Cinematic Conversations},
  author    = {tiger26},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/tiger26/cinebot-qwen3.5-9b}
}