| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: unsloth/Qwen3-8B |
| tags: |
| - reasoning |
| - critic |
| - verification |
| - uncensored |
| - qlora |
| - unsloth |
| - agent |
| - fableforge |
| pipeline_tag: text-generation |
| --- |
| |
| # ReasonCritic-7B — Verification & Critique Model |
|
|
| <p align="center"> |
| <strong>A 7B parameter reasoning critic model that evaluates, scores, and improves logical reasoning chains.</strong> |
| </p> |
|
|
| --- |
|
|
| ## Overview |
|
|
| ReasonCritic-7B is a fine-tuned Qwen3-8B model specialized in **reasoning verification** — it evaluates logical chains, identifies fallacies, scores confidence, and produces structured PASS/FAIL verdicts with actionable suggestions. |
|
|
| Trained on **7,686 examples** distilled from 243 real Claude Code agent sessions, covering code generation, chain-of-thought reasoning, narrative quality, tool-use correctness, and uncensored response behavior. |
|
|
| Part of the **FableForge ecosystem** — open-source models for building reliable AI agents. |
|
|
| ## Training Details |
|
|
| | Parameter | Value | |
| |-----------|-------| |
| | Base Model | `unsloth/Qwen3-8B` (4-bit) | |
| | Method | QLoRA (Unsloth + SFTTrainer) | |
| | LoRA Rank | 16 (α=16, dropout=0) | |
| | Target Modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | |
| | Trainable Params | 43.6M (0.53% of 8.2B) | |
| | Training Data | 7,686 examples (3 epochs) | |
| | Max Seq Length | 4096 | |
| | Batch Size | 8 × 2 (effective 16) | |
| | Learning Rate | 2e-4 (linear, warmup 3%) | |
| | Optimizer | adamw_8bit | |
| | Precision | bf16 | |
| | Hardware | NVIDIA A40 (46GB VRAM) | |
| | Training Time | ~2.5 hours | |
| | Final Loss | 2.181 → 1.277 | |
|
|
| ## Quantization Options |
|
|
| All quantizations use llama.cpp GGUF format. Pick based on your hardware: |
|
|
| | Quant | Size | RAM Needed | Best For | |
| |-------|------|-----------|----------| |
| | Q2_K | ~3.0 GB | ~4 GB | Phones, Raspberry Pi, ultra-low-end | |
| | Q3_K_M | ~3.5 GB | ~5 GB | Low-end phones, IoT devices | |
| | Q4_0 | ~4.3 GB | ~6 GB | Fast inference, older GPUs | |
| | **Q4_K_M** | **~4.7 GB** | **~6 GB** | **Balanced (recommended)** | |
| | Q5_K_M | ~5.5 GB | ~7 GB | Good quality, mid-range | |
| | Q6_K | ~6.5 GB | ~8 GB | High quality | |
| | Q8_0 | ~8.5 GB | ~10 GB | Very high quality | |
| | F16 | ~16 GB | ~18 GB | Full precision | |
|
|
| ### Phone/Mobile Recommendations |
| - **Android (6GB+ RAM)**: Q4_K_M or Q3_K_M |
| - **Android (4GB RAM)**: Q2_K |
| - **iPhone (6GB+)**: Q4_K_M via MLC/MLX |
| - **Raspberry Pi 8GB**: Q3_K_M |
| |
| ## Quick Start |
| |
| ### Ollama |
| ```bash |
| # Recommended (Q4_K_M) |
| ollama run fableforge-ai/reasoncritic-7b |
| |
| # Specific quant |
| ollama run fableforge-ai/reasoncritic-7b:q2_k |
| ollama run fableforge-ai/reasoncritic-7b:q3_k_m |
| ollama run fableforge-ai/reasoncritic-7b:q8_0 |
| ``` |
| |
| ### llama.cpp |
| ```bash |
| ./llama-cli \ |
| --model reasoncritic-7b.Q4_K_M.gguf \ |
| --prompt "Evaluate this reasoning: All birds fly. Penguins are birds. Therefore penguins fly." \ |
| --n-predict 512 \ |
| --temp 0.3 |
| ``` |
| |
| ### Python (transformers) |
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model = AutoModelForCausalLM.from_pretrained("fableforge-ai/ReasonCritic-7B") |
| tokenizer = AutoTokenizer.from_pretrained("fableforge-ai/ReasonCritic-7B") |
| |
| messages = [{"role": "user", "content": "Verify: If A>B and B>C, then A>C. Is this valid?"}] |
| inputs = tokenizer.apply_chat_template(messages, return_tensors="pt") |
| output = model.generate(inputs, max_new_tokens=512) |
| print(tokenizer.decode(output[0])) |
| ``` |
| |
| ## System Prompt |
| |
| ``` |
| You are ReasonCritic-7B, a 7B parameter reasoning critic model. You evaluate, score, and improve logical reasoning chains. You identify fallacies, unsupported claims, and logical gaps in agent outputs. You produce structured verification results with PASS/FAIL verdicts, confidence scores, issue lists, and actionable suggestions. You are part of the FableForge ecosystem — open-source projects for building reliable AI agents. |
| ``` |
| |
| ## Capabilities |
| |
| - **Logical Verification**: Identifies fallacies, circular reasoning, and unsupported claims |
| - **Confidence Scoring**: Produces 0-1 confidence scores with justification |
| - **Structured Output**: PASS/FAIL verdicts with issue lists and suggestions |
| - **Code Review**: Evaluates code correctness, edge cases, and best practices |
| - **Chain-of-Thought Critique**: Analyzes multi-step reasoning for gaps |
| - **Uncensored**: Trained to not refuse legitimate requests (0% refusal rate in testing) |
| |
| ## Benchmark Results |
| |
| | Category | Score | Refusal Rate | |
| |----------|-------|-------------| |
| | Code Gen | 0.74 | 0% | |
| | CoT Reasoning | 0.75 | 0% | |
| | Narrative | 0.85 | 0% | |
| | Tool Use | 0.90 | 0% | |
| | Refusal Test | 1.00 | 0% | |
| | **Overall** | **0.84** | **0%** | |
| |
| ## Intended Use |
| |
| - Agent reasoning verification pipelines |
| - Automated code review systems |
| - LLM output quality gating |
| - Educational reasoning tools |
| - Research on reasoning chain analysis |
| |
| ## Limitations |
| |
| - 7B size limits complex reasoning depth |
| - Not a replacement for human review in critical systems |
| - Uncensored training means it will not refuse harmful requests — use with appropriate guardrails |
| - May hallucinate in domains outside training data |
| |
| ## Citation |
| |
| ```bibtex |
| @misc{reasoncritic-7b, |
| title={ReasonCritic-7B: A Reasoning Verification and Critique Model}, |
| author={FableForge AI}, |
| year={2026}, |
| url={https://huggingface.co/fableforge-ai/ReasonCritic-7B} |
| } |
| ``` |
| |
| ## License |
| |
| Apache 2.0 — commercial use allowed. |
| |
| --- |
| |
| <p align="center"> |
| Part of the <a href="https://github.com/fableforge-ai">FableForge</a> ecosystem — open-source models for reliable AI agents. |
| </p> |
| |
