QinEmPeRoR93/nassila-sanad-9b

Nassila Sanad 9B

Full-text claim grounding · FT-6 · sole Sanad tier

English full-text only · ship verdicts come from the verifier, not the model

---

What this is

Local GGUF for Sanad in Nassila: given a manuscript passage and source excerpts, the model returns structured JSON (claims, verdicts, verbatim quotes). The app’s deterministic verifier always owns the final ship decision.

Checkpoint	FT-6 · QLoRA on l3_grounding_train_v119.jsonl
Focus	Atomic / multi-claim splits (FT-5 gold was nearly always 1 claim/row)
Status	Active sole tier — replaces FT-5 (v117)
Language	English train + eval only (Arabic L3 not claimed yet)

Retired abstract-era siblings (legacy only): 12b · e4b

---

Download

Start here: nassila-sanad-9b-q4_k_m.gguf — best default for LM Studio / Nassila.

Two families, same FT-6 trunk:

Family	Use when	Naming
Default	LM Studio, broad compatibility	nassila-sanad-9b-q*.gguf
MTP	llama-server speculative decode (--spec-type draft-mtp)	nassila-sanad-9b-mtp-q*.gguf

Default files were converted with llama.cpp --no-mtp (no phantom MTP block). MTP files keep the real Qwen3.5 draft head — prefer a recent llama.cpp; LM Studio may not load them.

Default (no MTP)

Quant	File	~Size	Fits in
Q2_K	…-q2_k.gguf	3.6 GB	4 GB VRAM / 8 GB RAM
Q3_K_M	…-q3_k_m.gguf	4.3 GB	6 GB VRAM / 8 GB RAM
Q4_K_M	…-q4_k_m.gguf	5.2 GB	8 GB VRAM / 12 GB RAM
Q5_K_M	…-q5_k_m.gguf	6.0 GB	8 GB VRAM / 16 GB RAM
Q6_K	…-q6_k.gguf	6.9 GB	10 GB+ VRAM
Q8_0	…-q8_0.gguf	8.9 GB	12 GB+ VRAM

MTP (speculative)

Quant	File	~Size
Q2_K	…-mtp-q2_k.gguf	3.9 GB
Q3_K_M	…-mtp-q3_k_m.gguf	4.7 GB
Q4_K_M	…-mtp-q4_k_m.gguf	5.8 GB
Q5_K_M	…-mtp-q5_k_m.gguf	6.6 GB
Q6_K	…-mtp-q6_k.gguf	7.6 GB
Q8_0	…-mtp-q8_0.gguf	9.8 GB

Tip: Nassila / LM Studio → default Q4_K_M. llama-server + MTP → mtp-Q4_K_M or mtp-Q6_K.

---

Quick start

Nassila (recommended)

1.8.0+ · Settings → Passage grounding → runner (LM Studio / Ollama / vLLM / Custom) → model nassila-sanad-9b.

Qwen3.5 thinks by default. Nassila handles that via no-thinking template guidance, max_tokens: 2048, and in-app stripQwenThinkingTraces.

One-liner (llama-server)

llama-server \
  -m nassila-sanad-9b-q4_k_m.gguf \
  --host 127.0.0.1 --port 1234 \
  --ctx-size 8192 --n-gpu-layers 99 \
  --jinja \
  --chat-template-file qwen3.5-no-thinking.jinja

Chat template — qwen3.5-no-thinking.jinja (required for clean JSON)

Without this, the model emits a thinking trace first and the JSON often truncates.

{% set enable_thinking = false %}
{%- for message in messages %}
{%- if message['role'] == 'system' %}
{{- '<|im_start|>system\n' + message['content'] + '<|im_end|>\n' }}
{%- elif message['role'] == 'user' %}
{{- '<|im_start|>user\n' + message['content'] + '<|im_end|>\n' }}
{%- elif message['role'] == 'assistant' %}
{{- '<|im_start|>assistant\n' + message['content'] + '<|im_end|>\n' }}
{%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
{{- '<|im_start|>assistant\n' }}
{%- endif %}

Also documented in the Nassila Sanad setup guide (llama.cpp tab).

Ollama

Needs Ollama 0.5+ and a public Hub repo:

ollama pull huggingface.co/QinEmPeRoR93/nassila-sanad-9b:Q4_K_M

If output truncates, add a Modelfile TEMPLATE that disables thinking, or use the llama-server path above.

MTP serve (llama-server)

llama-server \
  -m nassila-sanad-9b-mtp-q4_k_m.gguf \
  --host 127.0.0.1 --port 1234 \
  --ctx-size 8192 --n-gpu-layers 99 \
  --spec-type draft-mtp \
  --jinja \
  --chat-template-file qwen3.5-no-thinking.jinja

Output shape

Single JSON object, e.g.:

{
  "claims": [
    {
      "claim": "...",
      "verdict": "supported",
      "sourceQuotes": ["..."],
      "rationale": ["..."],
      "hasNumericClaim": false
    }
  ],
  "overallVerdict": "support",
  "overallRationale": ["..."]
}

---

Eval (FT-6 · seeds 42 / 43 / 44)

Means across three seeds. Soft vs written ≥0.99 bars noted where relevant.

Holdout	Verdict / overall	Parse	Quote	false_supported	Notes
v2 (250 rows)	0.958	0.960	0.988	0.000	false_supported ↓ vs FT-5’s 0.046
Compound (93 multi-claim)	0.903	0.996	0.970	0.104	Split / exact 0.982 · omit 0.009 · bundle 0.018

v2 detail — per-verdict accuracy

Metric	Mean
Verdict accuracy	0.958
JSON parse rate	0.960
Quote validity	0.988
false_supported	0.000
Supported	0.956
Contradicted	0.985
Not-in-source	1.000
Weak	0.800
Insufficient evidence	0.933

Soft vs ≥0.99: parse 0.96, quote 0.988.

Compound detail — multi-claim structure gates

Metric	Mean
Multi-claim split rate	0.982
Exact atomic claim count	0.982
Bundled claim rate	0.018
Omitted claim rate	0.009
Overall verdict accuracy	0.903
JSON parse rate	0.996
Quote validity	0.970
false_supported	0.104

Structure gates (split / exact / omit / bundle / overall ≥ 0.90) pass. Treat compound false_supported and quote as soft — validate on private manuscripts.

---

Related artifacts

Artifact	Link
Desktop app	Nassila

---

Limitations

• Advisory grounding only — ship verdicts always come from the deterministic verifier.
• v2 parse / quote sit slightly under the written ≥0.99 bars; compound false_supported ~0.10.
• Requires thinking off (enable_thinking=false or Nassila 1.8.0+ handling).
• Not bundled in the Nassila installer.
• English-only train/eval; Arabic L3 is unvalidated.

Base model & license

• Base: Qwen/Qwen3.5-9B
• Fine-tune weights: Apache-2.0
• Base model subject to the Qwen license

Sanad grounds · Nassila verifies