Figo-25M-SFT-DPO

Figo-25M-SFT-DPO

Custom language model trained completely from scratch! It is a tiny, highly responsive model designed to follow conversational instructions.

Model Details

• Architecture: 25 Million Parameter Custom LLM
• Tokenizer: 10K Vocabulary Byte-Pair Encoding (BPE)
• Base Pre-training: Trained on the TinyStories dataset to learn basic English syntax and narrative structures.
• Alignment (SFT): Fine-tuned via Supervised Fine-Tuning to shift behavior from open-ended storytelling to an interactive assistant persona.
• Preference Optimization (DPO): Optimized via Direct Preference Optimization to stabilize chat formatting, and favor clear, concise outputs.

Try it live (runs entirely in your browser, no server): habiz/figo-25m-chat

Performance

Measured directly against the published weights (model.safetensors), reconstructed architecture, and the chat format <|system|>/<|user|>/<|assistant|>/<|end|>. Default sampling: temperature 0.8, top-k 40.

Language modeling quality

Cross-entropy loss / perplexity on 500 held-out examples (~104K tokens) from the roneneldan/TinyStories validation split (teacher-forced, no sampling):

Chat generation robustness

80 sampled generations (10 varied prompts × 8 seeds, temp 0.8, top-k 40, max 80 new tokens):

As expected for a 25M-parameter model, quality varies a lot by prompt, casual greetings terminate cleanly almost every time (100%), while arithmetic ("What is 2 + 2?") degenerates into repeated-digit loops more often than not (62% degeneration rate). This model is tuned for TinyStories-style conversational tone, not factual accuracy or math.

Inference speed

Single-batch, greedy decoding, 100 new tokens, full-context recompute (no KV cache), best-of-3:

The in-browser number is what visitors to the Space actually experience, single-threaded WASM is expectedly the slowest backend, but still fast enough for interactive chat given the model's size.