You can deploy this model and start earning money today!

piyushptiwari/
ebrm-v2-qwen3-4b

Name: ebrm-v2-qwen3-4b
Availability: InStock

Views

No views yet

[!WARNING] Superseded by piyushptiwari/ebrm-v4-qwen3-4b (2026-05). v4 was trained on top of v2 with hard-negative mining and beats v2/v3 by +2.8 pp on BBH best-of-6. Neither v2 nor v4 beats self-consistency majority voting at k=6 — see the v4 model card for an honest benchmark table. v2 is retained for reproducibility of earlier results.

EBRM v2 — Qwen3-4B

Energy-Based Reasoning Machine v2: scoring heads trained on top of Qwen/Qwen3-4B for self-consistency reasoning. Used together with the ebrm-system pipeline (intent routing → Langevin candidate generation → SymPy/Lean/exec verifiers → voting).

Files

File	Purpose
`ebrm_state.pt`	Full training checkpoint (pooler, projector, energy_head, answer_decoder, optimizer, EMA)
`ebrm_inference.pt`	Inference-only state (heads + EMA, no optimizer) — load this for inference
`training_log.json`	Per-step training metrics

The ebrm_state.pt includes the optimizer state (≈ half the file size) and is intended for resuming training. For inference, use ebrm_inference.pt.

Architecture

The four heads sit on top of a frozen Qwen/Qwen3-4B backbone:

Qwen3-4B hidden states  ──►  pooler (mean over tokens)
                              │
                              ▼
                          projector (MLP → latent z)
                              │
              ┌───────────────┴───────────────┐
              ▼                               ▼
         energy_head (z → E)         answer_decoder (z → tokens)

energy_head is the scoring function used by the EBRM voter; answer_decoder reconstructs an answer from the latent for self-consistency.

Quickstart (no GPU server required)

pip install ebrm-system transformers torch huggingface_hub

python

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer

# 1. Download the inference checkpoint
ckpt_path = hf_hub_download("piyushptiwari/ebrm-v2-qwen3-4b", "ebrm_inference.pt")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)

# 2. Load the base LM (frozen during EBRM training)
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
lm = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B", dtype=torch.float16, device_map="auto")

# 3. Use the ebrm-system pipeline
from ebrm_system.intent import Intent
from ebrm_system.verifiers.routing import chain_for_intent

chain = chain_for_intent(Intent.MATH_REASONING)
# ... generate candidates with `lm`, score with energy_head, verify with chain, vote.

For an end-to-end demo (KV-cache compression + 32-candidate self-consistency on GSM8K): see scripts/demo_e2e.py in the ebrm-system repository.

Hardware

Inference fits on a single 16 GB GPU (Tesla T4 verified). ebrm-system also supports CPU inference for small batches.

License

Apache-2.0. Base model: Qwen/Qwen3-4B (Apache-2.0).

Citation

bibtex

@software{ebrm_system_2026,
  author  = {Tiwari, Piyush},
  title   = {ebrm-system: Energy-Based Reasoning Machine pipeline},
  year    = {2026},
  url     = {https://github.com/piyushptiwari1/ebrm-system}
}