rjz123/colar-synth-r1q

CoLaR — Synthetic arithmetic (controlled variable-chains), R1-Distill-Qwen-1.5B

A CoLaR (Compressed Latent Reasoning) checkpoint for Synthetic arithmetic (controlled variable-chain reasoning), fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. The model reasons in compressed continuous latent embeddings rather than explicit chain-of-thought tokens.

Model

• Base model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
• Framework: CoLaR — Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (arXiv:2505.16552).
• Domain: Synthetic arithmetic (controlled variable-chain reasoning)
• Warm-start: base CoLaR
• Files: colar_ctrl_r1.ckpt, hparams.yaml

Training procedure

Trained with the CoLaR supervised fine-tuning (SFT) recipe: the frozen base LLM is adapted with q/v LoRA (rank 128, alpha 32) plus a trainable Latent Head (3-layer MLP) and an embedding-compression module. The objective is next-token cross-entropy on the answer plus an embed_modeling_loss (MSE) that reconstructs the compressed reasoning-step embeddings (each latent token summarizes about compression_factor chain-of-thought tokens). Reinforcement learning (GRPO) is disabled — this is an SFT-only checkpoint.

• This checkpoint: CoLaR SFT, compression factor 4, 50 epochs; SFT only. R1 base.

Datasets

• Synthetic controlled arithmetic (variable-chain)

How to load

This is a PyTorch-Lightning checkpoint (weights under the top-level key state_dict) that fits the CoLaR scaffold — it is not directly AutoModel-loadable. Load the base model, splice this state_dict in with strict=False, and use the CoLaR runtime settings:

COLAR_EMB_STD=0.018   COLAR_COMPRESS=<compression_factor>   sep_token=###
TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1

See the CoLaR repository (github.com/xiaomi-research/colar) for the exact loader. The shared GSM8K warm-start ancestor is the official AlbertTan/CoLaR release.

Research artifact for latent-reasoning study (small 1–1.5B model).