minicpm5
Datasets
All datasets matching “minicpm5”Transmem_ecsd_minicpm5_1b_hotpotqa_n4_n8MiniCPM5-1B-atlas
juiceb0xc0de/MiniCPM5-1B-atlas
A brain atlas for openbmb/MiniCPM5-1B, a 1B on-device model with a 130k bilingual vocabulary. This is not a chat dataset or a benchmark. It is an internal-mechanics map, built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.
If you want to know which parts of this model are safe to edit, where its output-vocabulary directions live, or which layers are carrying the most… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/MiniCPM5-1B-atlas.minicpm5-1b-SAEOne JumpReLU SAE per layer of MiniCPM5-1B. All 24 layers, complete.
MiniCPM5-1B: 24 layers, 1536-dim residual stream, 130,560-token bilingual vocab.
Every SAE in this repo: d_in=1536, 49,152 features (32x expansion), JumpReLU activation, streamed FineWeb-Edu, target sparsity L0=50. Same settings on every layer, no hyperparameter changes were applied in the run.
Each layer_NN_s0/ holds:
sae.pt - the weights
meta.json - config and final metrics
checkpoint_full.pt - full optimizer state… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/minicpm5-1b-SAE.latent-state-tracking-minicpm5
Tracking and Intervening on Latent State Dynamics in a Small Language Agent (MiniCPM5-2B)
Date: 2026-09-17
Model studied: openbmb/MiniCPM5-2B (2.52B params, 42 layers, hidden dim 2048)
Hardware: single RTX 3070 Ti (8GB) — all experiments run on consumer-grade hardware
Summary
We ask whether a small (2.5B-parameter) language model's hidden-state trajectory during generation contains a stable, low-dimensional structure that (a) is linearly decodable into task type… See the full description on the dataset page: https://huggingface.co/datasets/B2J/latent-state-tracking-minicpm5.minicpm5-stock-v2-forward-return
MiniCPM5 Stock v2 — Forward-Return Labels
Binary BUY/SELL stock-direction dataset where labels come from actual forward
5-day returns (BUY > +2%, SELL < -2%, middle band dropped), not news sentiment.
All features are strictly causal (no look-ahead): last 20 daily returns, RSI(14),
volume ratio vs 20d MA, 20d volatility, 5d/20d momentum, 20d relative strength vs SPY.
train_minicpm5_v2.jsonl — 5,056 rows, 16 tickers, class-balanced
val_minicpm5_v2.jsonl — 1,586 rows, 4 held-out… See the full description on the dataset page: https://huggingface.co/datasets/ewin-reg/minicpm5-stock-v2-forward-return.minicpm5-2b-damage-labels
MiniCPM5-2B Damage Labels (MERNIK teacher)
Per-group measured quantization damage for MiniCPM5-2B (dense 2.6B, 42 layers).
What
damage_minicpm5_2b.jsonl — 169 rows: 1 BASELINE + 168 tied-group units.
Each unit row: the group dropped Q5_K → Q3_K while everything else stays at
Q5_K, scored by wikitext-2 PPL (-c 1024 -n 64 --seed 7).
{"unit": "ffn_down@7", "tensors": ["blk.7.ffn_down.weight"],
"ppl": 13.5364, "damage": 0.1732}
ssim_minicpm.npz — measured structural… See the full description on the dataset page: https://huggingface.co/datasets/wepiqx/minicpm5-2b-damage-labels.