| --- |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| base_model: Qwen/Qwen3-14B |
| pipeline_tag: text-generation |
| library_name: transformers |
| tags: |
| - philosophy |
| - dpo |
| - causal-reasoning |
| - specialist-model |
| - tunedai |
| inference: false |
| --- |
| |
| # Philosopher-14B |
|
|
| A 14B parameter philosophical reasoning specialist, fine-tuned from `Qwen/Qwen3-14B` via Direct Preference Optimization (DPO) on curated philosophy preference data. |
|
|
| **Built by [TunedAI Labs](https://tunedailabs.com).** Demonstrates that domain-specialized small models can outperform general-purpose frontier models on their target domain — at roughly **1/17 the parameters of Qwen3-235B**. |
|
|
| --- |
|
|
| ## What this model is for |
|
|
| Philosopher-14B is built for **depth and thoroughness on philosophical questions** — covering the major positions, thinkers, dates, and works for any topic in the field, then going deeper on the underlying disagreements. |
|
|
| Where most chat models give a polished surface-level summary of a philosophical question, Philosopher-14B is trained to: |
|
|
| 1. Lay out the full landscape — every relevant theory, position, and thinker, including dates and major works |
| 2. Go deeper on each position — the underlying assumptions, where each argument holds and breaks |
| 3. Surface the root disagreement — what's actually at stake beneath the surface positions, what can be established, and what remains genuinely open |
|
|
| It is **not** a general-purpose assistant. It is a specialist tool for studying philosophy — useful for students, researchers, philosophy-curious readers, and anyone wanting deeper-than-Wikipedia treatment of a question. |
|
|
| ## Quick example |
|
|
| > **Q:** "Could there be a fact that is true but permanently unknowable?" |
|
|
| A general-purpose frontier model produces a summary of Fitch's paradox plus one or two passes of agnosticism. Philosopher-14B opens with Fitch, then walks through verificationism (Schlick, Carnap), epistemic constructivism (Dummett, Wright), Williamson on knowability, then goes back through each at depth — what each position actually rules out, where the moves break, and what survives. The depth is the point. |
|
|
| ## How to use it |
|
|
| ```python |
| from transformers import AutoTokenizer, AutoModelForCausalLM |
| from peft import PeftModel |
| import torch |
| |
| base_model = AutoModelForCausalLM.from_pretrained( |
| "Qwen/Qwen3-14B", |
| torch_dtype=torch.bfloat16, |
| device_map="auto", |
| ) |
| tokenizer = AutoTokenizer.from_pretrained("tunedailabs/philosopher-14b") |
| model = PeftModel.from_pretrained(base_model, "tunedailabs/philosopher-14b") |
| model.eval() |
| |
| messages = [{"role": "user", "content": "Do we have free will?"}] |
| text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
| inputs = tokenizer(text, return_tensors="pt").to(model.device) |
| |
| with torch.no_grad(): |
| output = model.generate(**inputs, max_new_tokens=2000, temperature=0.7, do_sample=True) |
| print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)) |
| ``` |
|
|
| **Recommended generation settings:** `temperature=0.7`, `max_new_tokens=2000–4000` for depth. Lower temperature (0.3–0.5) for more focused outputs. |
|
|
| **Note on `<think>` tokens:** Qwen3-14B's reasoning mode is preserved. Responses begin with a `<think>...</think>` block showing the model's internal reasoning trace. We left this on intentionally — for philosophy, the *how* of the reasoning is part of the value. If you prefer answers without thinking blocks, pass `enable_thinking=False` to the chat template. |
|
|
| ## Model details |
|
|
| | | | |
| |---|---| |
| | **Base model** | `Qwen/Qwen3-14B` | |
| | **Training method** | Direct Preference Optimization (DPO) | |
| | **Adapter type** | LoRA (provided as PEFT adapters; merge optional) | |
| | **Parameters** | 14B (base) + small LoRA delta | |
| | **Precision** | bf16 | |
| | **Context length** | 32K tokens (inherited from base) | |
| | **Languages** | English (training data was English-only) | |
|
|
| The training pipeline is part of TunedAI Labs' proprietary methodology and is not described in detail here. **What's released is the model behavior**; what's retained is the methodology that produced it. |
|
|
| ## Comparison: depth-over-size |
|
|
| A 14B specialist vs. a 235B generalist on the same prompt is the headline result. In our internal evaluations, Philosopher-14B produces measurably more thorough, deeper, and more historically-grounded answers on philosophical questions than `Qwen/Qwen3-235B-A22B-Instruct-2507` — a frontier-class generalist roughly 17× the parameter count. |
|
|
| **A live side-by-side comparison demo is available at:** |
| [`https://tunedailabs-philosopher-demo.hf.space`](https://tunedailabs-philosopher-demo.hf.space) (private demo, contact for access) |
|
|
| We're releasing the weights so others can verify this directly. We invite the community to construct formal philosophical-depth evaluations against this and other specialist models — we believe small + specialized > large + general for any well-bounded knowledge domain, and this model is offered as one piece of evidence. |
|
|
| ## What this model is *not* good at |
|
|
| - General-purpose chat (use a general model) |
| - Coding (use a coder model) |
| - Math, science, current events (use a frontier model) |
| - Recent history (knowledge cutoff matches base model) |
| - Languages other than English |
|
|
| We trained for depth on a single domain. That depth comes at the cost of breadth. |
|
|
| ## Limitations & caveats |
|
|
| - **Hallucinations.** Like all LLMs, this model can confidently produce incorrect philosophical claims, misattributions, or fabricated quotations. Verify any claim before citing in academic work. |
| - **Western-canon bias.** Training data over-represents Western analytic and continental traditions; non-Western philosophy is included but less deeply. |
| - **Not a replacement for primary texts.** Use this model to orient yourself in a question, not as a substitute for reading Hume / Kant / Wittgenstein / etc. directly. |
| - **No moral or psychological claims about the user.** This is an academic-philosophy tool, not a counselor. |
|
|
| ## License |
|
|
| **Creative Commons Attribution-NonCommercial 4.0 (CC-BY-NC-4.0).** |
|
|
| You may use, modify, and redistribute this model for **research, education, and personal use**. **Commercial use requires a separate license from TunedAI Labs** — contact us at hello@tunedailabs.com. |
|
|
| The base model `Qwen/Qwen3-14B` is licensed separately by Alibaba; please respect their terms in addition to this license. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{philosopher14b2026, |
| title = {Philosopher-14B: A Domain-Specialized Small Model for Philosophical Reasoning}, |
| author = {Mark Gentry and TunedAI Labs}, |
| year = {2026}, |
| url = {https://huggingface.co/tunedailabs/philosopher-14b}, |
| note = {Fine-tuned via DPO from Qwen3-14B} |
| } |
| ``` |
|
|
| ## Acknowledgments |
|
|
| Built on `Qwen/Qwen3-14B` from the Qwen team at Alibaba. Training infrastructure on [Modal](https://modal.com). |
|
|
| ## Contact |
|
|
| - **Website:** [tunedailabs.com](https://tunedailabs.com) |
| - **Email:** hello@tunedailabs.com |
| - **Live demo:** [tunedailabs-philosopher-demo.hf.space](https://tunedailabs-philosopher-demo.hf.space) |
| - **Author:** Mark Gentry, founder TunedAI Labs |
|
|
| --- |
|
|
| *Philosopher-14B is the first public release in TunedAI Labs' specialist-model program. We're applying the same methodology to causal reasoning across other domains — civil rights, intent detection, decision support. More results forthcoming.* |
|
|
