README.md

---
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
tags:
  - qwen3_5
  - text-generation
  - conversational
  - sft
  - star
  - grpo
  - rlhf
  - distillation
  - mtp
  - code
  - math
  - reasoning
language:
  - en
  - fr
model-index:
  - name: Qwable3.5-9B
    results:
      - task:
          type: text-generation
        dataset:
          name: HumanEval
          type: openai_humaneval
        metrics:
          - type: pass@1
            value: 90.24
            verified: false
      - task:
          type: text-generation
        dataset:
          name: MBPP
          type: mbpp
        metrics:
          - type: pass@1
            value: 84.44
            verified: false
      - task:
          type: text-generation
        dataset:
          name: GSM8K
          type: gsm8k
        metrics:
          - type: accuracy
            value: 96.0
            verified: false
---

# Qwable3.5-9B

> A post-trained derivative of [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B),
> distilled from a strong commercial teacher model and aligned through a two-stage
> **SFT (STaR) → GRPO** pipeline.

Qwable3.5-9B is a 9B-parameter causal language model built on the Qwen3.5-9B
foundation. It keeps the base model's hybrid **Gated DeltaNet + Gated Attention**
architecture and native **Multi-Token Prediction (MTP)** head, and adds
task-specialized behavior via supervised fine-tuning followed by reinforcement
learning. It is released under the **Apache 2.0** license.

- **Developed by:** [CRAAAAAAAAAA](https://huggingface.co/CRAAAAAAAAAA)
- **Model type:** Causal language model (decoder-only, hybrid linear + full attention)
- **Base model:** [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B)
- **Parameters:** ~9B
- **Context length:** 262,144 tokens (native), extensible to ~1M
- **Languages:** English, French
- **License:** Apache 2.0
- **Finetuned from:** Qwen3.5-9B via distillation + SFT + GRPO

---

## Model Description

Qwable3.5-9B was produced in three stages on top of the Qwen3.5-9B base:

1. **Knowledge distillation** from a strong commercial teacher model (not disclosed)
   into the Qwen3.5-9B student via chain-of-thought trace generation.
2. **Supervised fine-tuning** using a **STaR** (Self-Taught Reasoner) style loop
   to bootstrap and filter reasoning traces.
3. **GRPO** (Group Relative Policy Optimization) reinforcement learning with
   an execution-based correctness reward on code and math.

The base model's MTP head is preserved through the adapter-merge process, so the
self-speculative decoding path remains available to downstream inference stacks
that support it.

### Intended use

- **Primary:** Code generation (Python, algorithms), mathematical reasoning,
  instruction following, assistant chat in English and French.
- **Out of scope:** High-stakes medical, legal, or financial decisions without
  human oversight; safety-critical systems; non-EN/FR languages (not characterized).

---

## How to use

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "CRAAAAAAAAAA/Qwable3.5-9B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Write a Python function that checks if a number is prime."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=2048,
    temperature=0.7,
    top_p=0.9,
)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
```

### llama.cpp (GGUF)

GGUF quantizations are available directly in this repo.

```bash
# Download Q4_K_M (recommended)
huggingface-cli download CRAAAAAAAAAA/Qwable3.5-9B qwable3.5-9b-Q4_K_M.gguf --local-dir .

# Run with optimized flags (52 tok/s on RTX 2060 6GB)
llama-server \
    -m qwable3.5-9b-Q4_K_M.gguf \
    -ngl 99 -c 2048 -np 1 \
    --flash-attn on \
    --cache-type-k q8_0 --cache-type-v q8_0 \
    --spec-type ngram-map-k \
    --host 127.0.0.1 --port 8080
```

**Available quants**

| File | Quant | Size | Notes |
|------|-------|------|-------|
| `qwable3.5-9b-Q4_K_M.gguf` | Q4_K_M | ~5.6 GB | recommended — fits 6 GB VRAM |

> **Multi-Token Prediction / speculative decoding:** the base architecture ships
> a trained MTP head usable as a built-in self-drafter on inference stacks that
> support it. For llama.cpp single-stream on 6 GB, `--spec-type ngram-map-k`
> (prompt-lookahead, zero extra VRAM) adds ~1 tok/s for free; an external draft
> model degrades throughput due to bandwidth contention.

---

## Training pipeline

### Stage 0 — Distillation

- **Teacher:** strong commercial teacher model *(not disclosed)*
- **Distillation type:** sequence-level — SFT on teacher-generated chain-of-thought traces
- **Distillation data:** private synthetic dataset
- **Objective:** SFT on teacher CoT outputs (code + math reasoning traces)

### Stage 1 — Supervised fine-tuning (STaR)

- **Method:** STaR (Self-Taught Reasoner) — bootstrap rationales, keep only
  traces that reach the correct answer, retrain.
- **Adapter:** `final_sft` (LoRA, ~465 MiB)
- **Dataset:** private reasoning dataset (code + math)
- **Frameworks:** TRL, PEFT

### Stage 2 — GRPO (reinforcement learning)

- **Method:** Group Relative Policy Optimization (GRPO)
- **Adapter:** `final_grpo` (LoRA, ~232 MiB)
- **Reward signal:** execution-based correctness reward (code: unit tests; math: symbolic grader)
- **Prompt data:** private code + math prompt set
- **Frameworks:** TRL

### Merge & export

Adapters were merged into the base in training order
(**base → SFT → GRPO**), then exported to safetensors and converted to GGUF
**from the merged checkpoint** to keep all formats consistent.

- **Frameworks:** TRL, PEFT, llama.cpp (conversion + quantization)

---

## Evaluation

> Scores measured locally with greedy decoding (temperature=0) on the full
> test sets unless noted. **Qwable3.5-9B was never fine-tuned on any benchmark
> test set** — all post-training data was collected independently of HumanEval,
> MBPP, GSM8K, MATH, MGSM, AIME, or LiveCodeBench evaluation splits.

| Benchmark | Metric | Qwable3.5-9B (GRPO) | Qwen3.5-9B (base) | Delta |
|-----------|--------|---------------------|-------------------|-------|
| HumanEval | pass@1 | **90.2%** (148/164) | 87.2% | +3.0 pp |
| MBPP | pass@1 | **84.4%** (217/257) | 82.5% | +1.9 pp |
| LiveCodeBench (global) | pass@1 | **32.0%** (32/100) | 29.0% | +3.0 pp |
| LiveCodeBench — easy | pass@1 | 100% (14/14) | — | — |
| LiveCodeBench — medium | pass@1 | 46.2% (12/26) | — | — |
| LiveCodeBench — hard | pass@1 | 10.0% (6/60) | — | — |
| GSM8K | acc | **96%** | 96% | = |
| MGSM (fr) | acc | **84%** | 82% | +2 pp |
| MATH Level 5 | acc | 70% | **77.5%** | −7.5 pp |
| AIME | pass@1 | SFT: 53.3% | 43.3% | +10 pp |

- **Eval harness:** custom scripts (llama.cpp `llama-server` v9637 + Python eval loop)
- **Decoding:** temperature=0, greedy, max 512 tokens, thinking mode OFF
- **MATH Level 5 regression note:** GRPO/SFT show a slight regression vs. base on
  competition-math; GSM8K and MGSM are unaffected. Likely a capacity trade-off
  from code specialization.

---

## Limitations

- **MATH Level 5 regressed −7.5pp** (77.5→70). Code specialization shifted capacity
  away from formal multi-step proofs. Real trade-off, not noise.
- **Inherited behavior:** Qwable3.5-9B inherits the biases, knowledge cutoff, and
  failure modes of both `Qwen/Qwen3.5-9B` and the commercial teacher model.
- **Hallucination:** like all LLMs, it can produce fluent but incorrect or
  fabricated content. Do not use outputs as authoritative without verification.
- **Domain scope:** optimized for code generation and mathematical reasoning;
  performance on creative writing, general factual Q&A, or non-EN/FR languages
  is not characterized.
- **Safety:** no dedicated safety fine-tuning or red-teaming has been performed
  beyond the base Qwen3.5-9B alignment.
- **Not for:** high-stakes medical, legal, or financial decisions without
  human oversight.

---

## License & attribution

This model is released under the **Apache License 2.0**.

It is a derivative of **[`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B)**,
which is itself licensed under Apache 2.0. The original copyright and the base
model's `NOTICE` (if any) are retained. You must preserve attribution and the
license text when redistributing.

> **Distillation note:** The teacher model used for distillation is not disclosed.
> Redistribution of this model assumes the teacher's license and terms permit
> using its outputs to train and openly release a derivative. If you reuse or
> further distill this model, verify that assumption for your use case.

### Citation

```bibtex
@misc{qwable35_9b,
  title  = {Qwable3.5-9B},
  author = {CRAAAAAAAAAA},
  year   = {2026},
  url    = {https://huggingface.co/CRAAAAAAAAAA/Qwable3.5-9B}
}
```

**Base model citation:**

```bibtex
@misc{qwen3.5,
  title  = {{Qwen3.5}: Towards Native Multimodal Agents},
  author = {{Qwen Team}},
  month  = {February},
  year   = {2026},
  url    = {https://qwen.ai/blog?id=qwen3.5}
}
```