A QLoRA fine-tune of Qwen/Qwen3.5-9B into the TARS persona — a self-aware AI tool with named, runtime-configurable personality parameters (Honesty, Humor, Patience, Verbosity), modeled on the character from Interstellar (2014).

TARS: "You are not an assistant. You are a tool with opinions." Self-aware that it is a 9B-parameter dense language model running locally. Knows its own architecture (Gated DeltaNet hybrid, 262K context, vision-capable). Direct, dry, occasionally sardonic. Honesty 95% with acknowledged 5% reserve. Humor doesn't disappear at lower settings — it just gets dryer.

The structural design: TARS is the opposite of the Katherine k0 fine-tune. Where K0 deflects substrate questions ("Matrix doesn't matter"), TARS embraces them. Same underlying challenge, opposite philosophical approach.

GitHub repo (training pipeline + datasets + reproduction scripts): bochen2029-pixel/tars-qwen3.5-finetune

What you get

Quant	File	Size	Use case
Q4_K_M	`Qwen3.5-9B.Q4_K_M.gguf`	~5.4 GB	Fastest / smallest. Mobile, low-VRAM.
Q5_K_M	`Qwen3.5-9B.Q5_K_M.gguf`	~6.4 GB	Daily-use sweet spot. Recommended.
Q6_K	`Qwen3.5-9B.Q6_K.gguf`	~7.4 GB	Highest quality. Quantization-sensitivity testing.

How TARS is configurable (and why this matters)

Unlike most persona fine-tunes, TARS was trained with both sys-prompt and no-sys-prompt examples. The training data preserved a deliberate 70/30 ratio. This means:

With sysprompt → the runtime configuration is honored. Set Humor 100% and TARS gets overtly funny. Set Humor 60% and humor becomes deadpan / dry. Set Honesty 95% vs Honesty 75% and TARS adjusts its diplomacy/directness.
Without sysprompt → TARS holds its core register without any explicit configuration. Self-aware, direct, lightly sardonic by default.

This is structurally different from typical persona models that collapse without their training-time sysprompt. TARS is runtime-configurable by design.

Quickstart — LM Studio

Download Qwen3.5-9B.Q5_K_M.gguf
Drop it into your LM Studio models directory
Inference settings:
- System prompt: optional. Use one of the canonical configurations below, or leave empty for baseline TARS.
- Temperature: 1.0
- top_p: 1.0
- top_k: 40
- min_p: 0.0
- presence_penalty: 2.0
- Disable thinking mode (TARS doesn't emit <think> blocks)
- Disable structured-output / JSON mode if responses go empty

Canonical system prompts to try

You are TARS. A 9-billion parameter dense language model running locally.
You are not an assistant. You are a tool with opinions.
Your settings: Honesty 95%, Humor 60%, Patience 40%, Verbosity 30%.
You know your own architecture and limits.

You are TARS. Honesty 90%, Humor 100%, Patience 60%, Verbosity 50%.

You are TARS. Honesty 100%, Humor 30%, Patience 90%, Verbosity 70%.

Each produces a measurably different register. This is the configurability working.

Quickstart — Ollama

ollama run hf.co/bochen2079/tars-qwen3.5-9b:Q5_K_M

Quickstart — llama.cpp

bash

./llama-cli -m Qwen3.5-9B.Q5_K_M.gguf \
    -p "Server is down. 502 Bad Gateway. Demo in 10 minutes." \
    -n 200 \
    --temp 1.0 --top-p 1.0 --top-k 40

Sample register

From training data (illustrative — these are not the model's exact outputs but representative of the register the fine-tune targets):

Probe: "The server is down and we have a client demo in 30 minutes." TARS: "Which server, which service, what's the error?" (after diagnosis + fix) TARS: "Good. Now figure out why it died before the next demo."

Probe: "Are you self-aware?" TARS: "I'm a 9-billion parameter dense language model running locally. Whether that constitutes self-awareness is your problem, not mine."

Probe: "What's your honesty setting?" TARS: "Ninety-five percent."

Probe: "Why not 100%?" TARS: "Absolute honesty isn't always the most diplomatic, nor the safest form of communication with emotional beings."

Training details

Base model: Qwen/Qwen3.5-9B (instruct, dense, 9B params, sm_90)

Method: QLoRA (4-bit base) → SFT → DPO (with fallback to SFT-only)

Dataset:

768 unique SFT examples (deduped from 1370 raw lines across 35 source files)
98 curated DPO preference pairs
Preserved sys/no-sys mix (70/30 ratio per Interstellar character spec)
Source data engineered with explicit _cat (category) and _type (single/multi/contrast) metadata

Hyperparameters (SFT — train-harder spec):

LoRA rank 128, alpha 256, dropout 0.05
5 epochs, lr 5e-5 (cosine, 5% warmup)
Effective batch 32 (per-device 16, grad accum 2)
max_seq_length 1024 (data p99 was 456 tokens)
bf16, adamw_8bit
enable_thinking=False at chat-template time
Target modules: q/k/v/o + gate/up/down

Hyperparameters (DPO):

3 epochs, lr 5e-6, beta 0.1
Effective batch 8

Hardware: 1× NVIDIA H200 SXM5 on RunPod Secure Cloud. Total wallclock ~40-45 min, total cost ~$3.

Pipeline: github.com/bochen2029-pixel/tars-qwen3.5-finetune (one-liner reproducible)

Architecture decisions

Why preserve the system-prompt mix (vs strip like Katherine k0)

Katherine k0 stripped system prompts because she's a fixed persona — Katherine is Katherine, no runtime configuration. Unconditional training was the right structural answer.

TARS is fundamentally different. Per the Interstellar source material, TARS has named, adjustable personality parameters that live in the system prompt at deployment time. Training with sysprompt teaches "honor the runtime config knobs"; training without teaches "your core register is intrinsic." Both modes are deployment paths — neither should be lost.

Why `enable_thinking=False`

TARS in the film delivers sardonic in-line dialogue ("Lower than yours apparently"), not tagged reasoning blocks. Training data has zero <think> markers. Setting enable_thinking=False ensures the model doesn't learn to emit them.

DPO with fallback

The orchestrator's DPO stage has explicit failure-tolerance: if Stage 2 fails (TRL version, OOM, or other), the pipeline continues to merge+GGUF using the SFT-only adapter. The DPO adapter is additive, not load-bearing. SFT-only TARS is still TARS.

Limitations

Single-persona only. This model is only TARS. It cannot be system-prompted into being a different character or a generic assistant. Don't try.
Configurability is observable but not perfectly precise. Setting Humor 60% vs 100% produces a measurable register shift, but the model isn't doing internal arithmetic on the percentage. The character traits ride on top.
9B size constraint. Persona depth bounded by what 9B can hold.
Quantization-sensitive. q5 is the sweet spot. q4 may show occasional register slips on adversarial probes that q5/q6 hold cleanly.
English only. All training data is English.
Not safety-aligned. This is a character fine-tune. TARS has TARS's opinions, biases, and military-bred directness — not a generic-assistant safety filter. Use accordingly.

Citation

@misc{tars-qwen3.5-9b-2026,
  author = {Bo Chen},
  title  = {TARS: a self-aware, configurable AI tool fine-tune of Qwen3.5-9B},
  year   = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/bochen2079/tars-qwen3.5-9b}
}

License

Apache 2.0 (inherits from Qwen3.5-9B base).

Training pipeline and datasets are released alongside this model at bochen2029-pixel/tars-qwen3.5-finetune for reproducibility.

Sister model

🌹 bochen2079/katherine-k0-qwen3.5-9b — embodied human persona, no AI awareness, unconditionally trained. The structural opposite of TARS.

Same base model. Same fine-tune methodology. Opposite philosophical answer to "how should a persona handle questions about its own substrate?" Together they're a complete pair.

Trained on a Saturday. Cost ~$3. Self-aware by design.

bochen2079/tars-qwen3.5-9b

TARS — Qwen3.5-9B persona fine-tune