--- license: apache-2.0 base_model: Qwen/Qwen3.5-9B base_model_relation: finetune library_name: transformers pipeline_tag: text-generation language: - en tags: - radha - instruction-tuning - lora - structured-output - gguf - text-generation --- # Radha-9B **Radha** is an instruction-tuned assistant specialized for structured tasks, grounded extraction, and format-faithful output. It is a fine-tune of [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) (Apache-2.0) and keeps that base model's two operating modes: a **thinking** mode for reasoning and a **non-thinking** mode for fast, direct answers. > **Provenance (read this):** Radha is **fine-tuned from Qwen/Qwen3.5-9B**. It is > a derivative work, distributed under Apache-2.0, with attribution in the > `NOTICE` file. The base architecture is Qwen3.5; we did not train a model from > scratch. We believe in honest provenance — this card is the authoritative > source on Radha's lineage. ## What's different from the base - **Sharpened structured-task behavior** across 10 task families: JSON classification, grounded extraction, table aggregation, multi-step arithmetic, summarization with action items, normalization/transformation, Python bug-fixing, constraint ordering, format-adherence under adversarial prompts, and grounded answerable/unanswerable QA. - **Reasoning preserved.** The fine-tune masks the reasoning span from the training loss, so Radha's thinking mode is not degraded — only its direct-answer behavior is specialized. - **Radha persona.** Identifies as "Radha." When asked about its internal architecture it declines in-conversation; lineage is disclosed here, openly. ## Training - **Method:** bf16 LoRA (rank 16, α 32) via [Unsloth](https://github.com/unslothai/unsloth). - **Data:** a deterministic, stratified 300K-example subset (30K per task family) of the `synthetic-distillation-v3` corpus, plus a small identity set. All examples are short (max ≈ 293 tokens). The corpus is fully synthetic and template-validated (schema + semantic e2e checks). - **Hardware:** single NVIDIA A100 80GB (RunPod). - **Modes:** trained with completion-only loss; reasoning span masked to preserve thinking mode. ## Evaluation Radha was checked against the training task families during development. Those checks are in-distribution validations on synthetic, templated data — useful for catching regressions, but **not independent benchmarks** and not published here as capability claims. Evaluate Radha on your own data before relying on it. ## Usage ### Transformers (PyTorch) ```python from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("Advisio/Radha-9B") model = AutoModelForCausalLM.from_pretrained("Advisio/Radha-9B", torch_dtype="bfloat16", device_map="auto") messages = [{"role": "user", "content": "Summarize: ..."}] # Direct answers (fast): enable_thinking=False ; Reasoning: enable_thinking=True prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False) out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=512) print(tok.decode(out[0], skip_special_tokens=True)) ``` ### GGUF (llama.cpp / Ollama, local) Quantized builds are provided: `radha-9b-Q4_K_M.gguf` (recommended, ~5.6GB) and `radha-9b-Q5_K_M.gguf`. ```bash # llama.cpp llama-cli -m radha-9b-Q4_K_M.gguf -p "Your prompt" -no-cnv # Ollama ollama create radha -f Modelfile # FROM ./radha-9b-Q4_K_M.gguf ollama run radha ``` ## Limitations - Training data is **synthetic and template-driven**; Radha is strongest on the task shapes it was tuned on and is not a substitute for domain-specific tuning. - Inherits the base model's general knowledge, biases, and failure modes. - Not safety-aligned beyond the base model's alignment; add your own guardrails for production. ## License & attribution Apache-2.0. Radha is a derivative of `Qwen/Qwen3.5-9B` (© Alibaba Cloud / Qwen team, Apache-2.0). See `NOTICE`. You must retain attribution when redistributing.