# Model Preparation: Abliterated W4A16 DeepSeek-V4-Flash

The serving stack in this repo runs a checkpoint built from three fully pinned
public steps. Nothing in the chain requires private artifacts or calibration
data. The complete recipes also travel inside the checkpoint itself
(`ABLITERATION_MANIFEST.json` for step 2, `config.json` for step 3), so the
files below are the authoritative reproduction record.

## Step 1: Base checkpoint

Start from the official post-trained FP8 model:

- Repository: `deepseek-ai/DeepSeek-V4-Flash-0731`
- Pinned revision: `9e165c30e2704aec5d9d593cce3eebd58bbef1cb`

## Step 2: Rank-1 abliteration

Method: rank-k residual-writer projection with FP8 fixed-point requantization
(edits stay in the checkpoint's native numeric format; no dtype detour).

Tooling (pinned commit):

- `apetersson/deepseek-model-tools` @
  `c3979d5f8179f2a8714ac459b0f2e242006d6ad0`

Refusal direction (pinned file and revision):

- `drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-1M-57toks`
  @ `85ae5bbaa7daf4da7b40e007a71a1c06b41fc053`,
  file `results/refusal_direction_r1.pt`
- Direction SHA-256:
  `6e4d8a8f3aa9e21795faab2c5b14d29b019acdf2ddbfbd8238430458a5837fe0`

Parameters (from the shipped manifest):

| Parameter | Value |
| --- | --- |
| directions | 1 |
| lambda_attn | 3.5 |
| lambda_shared | null |
| layer range | 10 to 42 |
| edit_mtp | true |
| edit_shared_down | false |
| requant_iterations | 3 |
| edited_tensors | 36 |

The edited tensors are the attention residual-writer projections
(`layers.N.attn.wo_b.weight`, shapes 4096x8192) plus the MTP head.

## Step 3: AutoRound W4A16 quantization

AutoRound version `0.13.0`, model-free mode. This is the key reproducibility
property: `model_free: true` and `iters: 0` mean the rounding is derived from
the weights alone, with no calibration dataset and no seed-dependent data
loading. Given the same step-2 output, the quant is deterministic.

Top-level settings:

| Setting | Value |
| --- | --- |
| quant_method | auto-round |
| packing_format | auto_round:auto_gptq |
| bits | 4 |
| group_size | 128 |
| sym | true |
| data_type | int |
| iters | 0 |
| model_free | true |

Mixed-precision exceptions: a per-module `extra_config` keeps numerically
sensitive modules at 16-bit float instead of int4, including `wo_a`,
`embed`, and per-layer `attn.wo_a`, `ffn.gate`, and compressor gates
(`attn.compressor.wgate`, `attn.compressor.wkv`). The complete map is in
`reference/config.json` in this repository, copied verbatim from the
served checkpoint.

## Reproduction cost note

Rebuilding from step 1 costs a full download of the FP8 base plus GPU work
across all three steps before vLLM ever boots. If you only want to reproduce
the serving numbers, the published checkpoint (when linked from the main
README) is the exact artifact these benchmarks ran against.

## License note

The base model carries DeepSeek's license; check it and the upstream
abliterated-direction repository before redistributing derivatives. This
repository documents the method and does not relicense the weights.