Saffron-v1-chat (Experimental)

Instruction-tuned version of Abhilash-AI-Lab/saffron-v1 — a ~100M-parameter English model from the Abhilash AI Research Lab, fine-tuned to follow simple instructions.

Status: Experimental. Lightly instruction-tuned at ~100M parameters. It follows the chat format but has little world knowledge and will hallucinate (including about its own identity). Not production quality, not safety-tuned, English only.

Chat template

<|user|>
{your message}
<|assistant|>
{reply}<|endoftext|>

Training

• Base model: Abhilash-AI-Lab/saffron-v1 (~100M, custom architecture + byte-level BPE)
• SFT data: Databricks Dolly-15k + Alpaca, assistant-response-only loss
• Best validation loss: 2.408 (perplexity 11.11)

Usage

Raw PyTorch checkpoint. Use the code at https://github.com/shishodiaabhilash/saffron-v1 :

python -m src.sample --config configs/sft.yaml --hf-repo Abhilash-AI-Lab/saffron-v1-chat --prompt "Hello"
# then chat locally:
python -m src.chat   --config configs/sft.yaml --ckpt results/saffron.pt

Limitations

Preliminary research artifact. Short, often-inaccurate answers; no factual grounding. Do not rely on outputs.