You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

cyberviser/opus-4.8-recreation-1b-light

Opus 4.8 Recreation (student) via Claude 4.8 Distillation + OpenMythos RDT

This is a governance-hardened experimental student model trained as part of the "Opus 4.8 recreation" research track.

Key characteristics

Base architecture: OpenMythos Recurrent-Depth Transformer (Prelude → RecurrentBlock×T (Parcae LTI + MLA/MoE + ACT) → Coda)
Distillation objectives (when enabled): logit KL + recurrent state matching + ACT halt prob + MoE router KL from Claude 4.8 extended thinking traces
Continued pretraining: FineWeb-Edu (sample-10BT) streaming
Run config: 2000 steps | light=True | distillation every 100 steps | unroll=1
Provenance: Full audit via claude_teacher_recursive_trainer.governance_precheck, check_spend_limits, DISTILLATION_STATE_FILE, execution_lock.json

Governance & Safety

This artifact was produced under the lab threat model and execution lock:

I_APPROVE_CLAUDE_4_8_DISTILLATION approval phrase + opus_4_8_recreation_allowed flag required
Persistent spend tracking and hard pre-flight budget checks (CLAUDE_*_BUDGET envs)
REVIEW_REQUIRED evidence gate for high-spend runs
See lab/policies/active_lab_threat_model.json (case: "Opus 4.8 recreation + capability exfiltration")

Strong disclaimer: This is an independent, open research reconstruction. Not affiliated with, endorsed by, or sponsored by Anthropic. Capabilities are intentionally limited by the 1B-scale student + conservative unroll + governance caps.

Loading (matching cyberviser org convention)

import json, torch
from open_mythos import OpenMythos, MythosConfig
from huggingface_hub import hf_hub_download

cfg_path = hf_hub_download("cyberviser/opus-4.8-recreation-1b-light", "config.json")
wts_path = hf_hub_download("cyberviser/opus-4.8-recreation-1b-light", "pytorch_model.bin")

with open(cfg_path) as f:
    cfg = MythosConfig(**json.load(f))

model = OpenMythos(cfg)
model.load_state_dict(torch.load(wts_path, map_location="cpu", weights_only=True))
model.eval()

Files

config.json — MythosConfig
pytorch_model.bin — state_dict
model.safetensors — (if generated)
final_model.pt — full checkpoint with provenance (in the run volume)