Configuration Parsing Warning:Invalid JSON for config file config.json

m1b-sft-lc

A checkpoint from MambaRL: does outcome-only RL extend the effective context-retrieval horizon of a pure Mamba-2 LM, without changing the architecture? No attention is added and the recurrent state is not grown — only what the policy learns to keep in it changes.

Base model: AntonV/mamba2-2.7b-hf
LoRA adapter merged in from checkpoints/M1b_sft_lc (targets: in_proj only — PEFT refuses out_proj/conv1d on mamba2 because the fused kernels read those weights directly)
Weights here are merged, so AutoModelForCausalLM.from_pretrained is enough.

Phase 1 (proposal 2.4): generic CoT LoRA SFT on 65536 examples, final train loss 0.0849, 0.98 h on one L40.

Answers follow a <think>...</think><answer>...</answer> contract.

Downloads last month: 5

Safetensors

Model size

3B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Lixing-Li/m1b-sft-lc

Base model

AntonV/mamba2-2.7b-hf

Adapter

(4)

this model