README.md

---
tags:
- opsd
- qwen3.5
- model-registry
---

# OPSD model registry

This repository is the canonical public home for OPSD model artifacts owned by
`LSW142857`. Indexed Medium, Strong, and Weak trailing-user full-model
milestones include their complete weights directly in the corresponding
iteration directories.

## Layout

```text
OPSD/
├── strong/
│   ├── first_user/iterations/
│   └── trailing_user/iterations/
├── medium/
│   ├── first_user/iterations/
│   └── trailing_user/iterations/
└── weak/
    ├── first_user/iterations/
    └── trailing_user/iterations/
```

Each available `iter_XXXXXXX` directory contains model files and provenance.
The number is the zero-indexed training iteration, so `iter_0000031` means 32
completed optimizer updates. An arm-level README records whether an experiment
is available, still training, private, or not yet run.

## Available artifacts

| PI | Teacher prompt | Iteration | Updates | Format | Direct load | Folder |
|---|---|---:|---:|---|---|---|
| Medium | `trailing_user` | 3 | 4 | merged full model | yes | [open](medium/trailing_user/iterations/iter_0000003/) |
| Medium | `trailing_user` | 7 | 8 | merged full model | yes | [open](medium/trailing_user/iterations/iter_0000007/) |
| Medium | `trailing_user` | 15 | 16 | merged full model | yes | [open](medium/trailing_user/iterations/iter_0000015/) |
| Medium | `trailing_user` | 31 | 32 | merged full model | yes | [open](medium/trailing_user/iterations/iter_0000031/) |
| Strong | `trailing_user` | 3 | 4 | merged full model | yes | [open](strong/trailing_user/iterations/iter_0000003/) |
| Strong | `trailing_user` | 7 | 8 | merged full model | yes | [open](strong/trailing_user/iterations/iter_0000007/) |
| Strong | `trailing_user` | 15 | 16 | merged full model | yes | [open](strong/trailing_user/iterations/iter_0000015/) |
| Strong | `trailing_user` | 31 | 32 | merged full model | yes | [open](strong/trailing_user/iterations/iter_0000031/) |
| Weak | `trailing_user` | 3 | 4 | merged full model | yes | [open](weak/trailing_user/iterations/iter_0000003/) |
| Weak | `trailing_user` | 7 | 8 | merged full model | yes | [open](weak/trailing_user/iterations/iter_0000007/) |
| Weak | `trailing_user` | 15 | 16 | merged full model | yes | [open](weak/trailing_user/iterations/iter_0000015/) |
| Weak | `trailing_user` | 1 | 2 | unmerged LoRA + full-MTP bundle | no | [open](weak/trailing_user/iterations/iter_0000001/artifact.json) |
| Weak | `trailing_user` | 2 | 3 | unmerged LoRA + full-MTP bundle | no | [open](weak/trailing_user/iterations/iter_0000002/artifact.json) |
| Weak | `trailing_user` | 4 | 5 | unmerged LoRA + full-MTP bundle | no | [open](weak/trailing_user/iterations/iter_0000004/artifact.json) |

The milestone labels count completed optimizer updates. The directory labels
use zero-indexed iterations, so update 4/8/16/32 maps to iteration 3/7/15/31.
The migrated Weak branch does not contain `iter_0000000`, so no one-update Weak
model is published. Its first two cumulative updates used `first_user`; later
updates used `trailing_user`. See the Weak arm README and per-model manifests.

## Direct load from a subfolder

```python
from transformers import AutoModelForCausalLM, AutoProcessor

repo = "LSW142857/OPSD"
subfolder = "medium/trailing_user/iterations/iter_0000031"
processor = AutoProcessor.from_pretrained(
    repo, subfolder=subfolder, trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    subfolder=subfolder,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
```

Machine-readable inventory: [`registry.json`](registry.json). Artifact schema:
[`schema/artifact.schema.json`](schema/artifact.schema.json).

## Naming contract

- Use only `strong`, `medium`, or `weak` for PI level.
- Use the exact code-level prompt modes `first_user` and `trailing_user`.
- Use zero-indexed, seven-digit directories: `iter_0000008`, not `checkpoint8`.
- Record the immutable Hugging Face commit in every `artifact.json`.
- Mark whether the artifact is directly loadable. Never describe an OPSD
  trainables bundle as a merged model.
- Do not place optimizer state, datasets, rollout transcripts, logs, or tokens
  in this public registry.

## Legacy repositories

Older public repositories whose prompt mode or artifact contract is not
unambiguously encoded are listed in
[`legacy_unclassified/README.md`](legacy_unclassified/README.md). They are not
silently assigned to an arm.