OpenMythos-2B β€” Edge (SFT + GRPO)

Qwen3.5-2B weights running inside the OpenMythos architecture (hybrid linear attention + GQA + MoE-ready blocks), converted 1:1, then instruction-tuned on UltraChat-200k and reinforced with GRPO using a LiquidAI/d1-3B decision-model verifier.

Base model Qwen/Qwen3.5-2B
Params 2.393 B (bf16)
Layers 24 (18 linear-attention + 6 full-attention), 11 + 1 recurrent + 12
Attention hybrid: Gated DeltaNet (linear) + GQA full-attention
Context 248 320 vocab, RoPE ΞΈ=1e7, 256k native positions
Weights 5.80 GB, 4 safetensors shards

πŸ“Š Verified results (Tesla T4, bf16)

Chat perplexity β€” identical held-out set: 69 UltraChat-200k test samples + 5 general-text samples = 12 313 tokens, same tokenizer and code path:

model held-out ppl vs Qwen
Qwen/Qwen3.5-2B (original) 7.9467 1.00Γ—
OpenMythos-2B (converted, pre-SFT) 11.8564 1.49Γ—
OpenMythos-2B (SFT β€” this repo) 6.2327 0.78Γ—

Math reasoning (GSM8K test, greedy generation):

model accuracy
OpenMythos-2B (converted, pre-SFT) 4 % (1/24)
OpenMythos-2B (after distillation) 29.2 % (7/24)

The conversion alone cost 1.49Γ— perplexity; 150 steps of distillation from a Qwen3.5-2B teacher closed the gap and overshot.

πŸ“¦ Files

file what
model-00001..00004-of-00004.safetensors the model (5.80 GB, bf16, 335 tensors)
model.safetensors.index.json shard index
config.json architecture metadata + training hyperparameters
sft2_distill_step150.safetensors the distillation checkpoint (29.2 % GSM8K)
grpo_step12.safetensors after GRPO with the d1-3B verifier
scratch_2b/ earlier from-scratch checkpoints (different, smaller architecture) β€” see its README

Load with the OpenMythos architecture (see config.json for the layer map); the weights are a 1:1 copy of the converted Qwen3.5-2B tensors.

πŸ—οΈ Architecture

Embed β†’ [11 Γ— TransformerBlock] β†’ [RecurrentBlock] β†’ [12 Γ— TransformerBlock] β†’ RMSNorm β†’ tied Head
                  β”‚                    β”‚
                  β”‚                    └─ ACT / LoRA recurrent reasoning block
                  └─ hybrid attention: Gated DeltaNet + GQA full-attention

Each TransformerBlock = pre-norm attention + SwiGLU feed-forward, with optional Mixture-of-Experts, Mixture-of-Depths routing, Multi-Token-Prediction head, sparse-attention (DSA) and long-term KV memory available as fine-tuning targets.

βœ… Conversion fidelity

Component Transfer
Gated DeltaNet attention (18 layers) direct copy β€” module parity 0.43 %
GQA full attention (6 layers) direct copy β€” layer parity cos 0.996
SwiGLU feed-forward (6144) direct copy
Norms (zero-centred RMSNorm) direct copy
Embedding / output head direct copy (tied)

load_state_dict on the promoted checkpoint reports 0 missing / 0 unexpected tensors.

Post-hoc compression of K and V into a shared MLA latent was measured too (SVD energy retention 86–98 %) and found lossy, so the converted checkpoint keeps per-head K/V; the compressed-latent attention path is kept in the project for training from scratch.

πŸ§ͺ Weight-Space Programming (gradient-free editing)

Weights can be edited directly, without gradients, as in Exploring Weight-Space Programming (arXiv:2611.09382). The admissible edit space is transformer-block parameters only (ffn.down.weight, attn.o_proj.weight).

The paper's headline edit was replicated on this model: contrasting the "reconsideration" discourse (Wait, However, But, Maybe, Perhaps) against the "forward-progress" discourse (Therefore, Thus, So, Now, The, First, Next, Let) in token-embedding space, then subtracting that direction from a single MLP column:

u = N( mean_A E - mean_B E ) * (1 + gamma)      gamma = final RMSNorm weight
W[:, j] -= u                                      only one column changes
weights changed 2 046 bf16 (0.0163 %) β€” the paper changed 1 023 (0.016 %)
target coda.11.ffn.down.weight column 869
dev NLL 3.0670 β†’ 3.0564
held-out NLL 3.0452 β†’ 3.0350

A control edit with the opposite sign regressed both dev and held-out, confirming the direction is meaningful rather than noise.

Edits are validated on a development split and only then measured on a held-out split β€” the same discipline the paper found necessary, where 6 of 16 trajectories had to be reverted because gains did not transfer.

πŸ“œ Attribution

Weights are derived from Qwen3.5-2B (Apache-2.0). Please respect the upstream license and keep the attribution when redistributing.

Downloads last month
1,356
Safetensors
Model size
3B params
Tensor type
BF16
Β·
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for okolxc/openmythos-2b-edge

Finetuned
Qwen/Qwen3.5-2B
Finetuned
(719)
this model

Paper for okolxc/openmythos-2b-edge