---
license: apache-2.0
tags:
- dflash
- speculative-decoding
- llava
---

# llava-v1.6-vicuna-7b-dflash-vlm-clean-overfit

A DFlash speculative-decoding draft model for
[`liuhaotian/llava-v1.6-vicuna-7b`](https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b),
trained on a small (50-row) image+question dataset as a **pipeline
proof-of-concept**, not a production drafter. Part of the
[VL-FLASH](https://github.com/arunmadhusud/VL-FLASH) project's Phase 5
(LLaVA-1.6 port).

## What this is

- Architecture: `DFlashDraftModel`, 5 layers, `hidden_size=4096`,
  `block_size=16` (config: `configs/llava-v1.6-vicuna-7b-dflash.json` in
  the [training fork](https://github.com/arunmadhusud/SpecForge/tree/feat/dflash-llava16)).
- Trained on 50 MMStar rows (real images, target-greedy regenerated
  answers), early-stopped at step 275/375 once pooled training accuracy
  held sustained above 90%.
- Warm-start: **none** — trained from scratch (unlike the sibling
  text-pretrain checkpoint in this same account, which this checkpoint
  predates).

## Measured result (cold-start replay gate, corrected per-row target)

Per-row target = `min(answer_token_len - 1, block_size - 1)` — LLaVA's
real VQA answers are often a single MCQ letter, so `accept_len` is
capped by each row's own answer length, not by `block_size`.

- **Short answers (< 16 tokens, n=26): 26/26 perfect.** Every single
  short answer (including every single-MCQ-letter row) replays at
  exactly its maximum possible `accept_len`.
- **Long answers (>= 16 tokens, n=24): 8/24 perfect.** The remaining 16
  were root-caused to a data-generation bug (`--max-tokens 60`
  truncating some training answers mid-generation) — fixed in the
  pipeline after this checkpoint was trained, not re-trained against the
  fix. See the repo's NOTES.md "Export + cold-start replay gate" section
  for the full breakdown.

## Reproduce / go further

Full runbook (this checkpoint's exact training run, plus how to scale up
to a bigger dataset and a bigger GPU):
[`scripts/gates/llava_text_only/RUNBOOK.md`](https://github.com/arunmadhusud/SpecForge/blob/feat/dflash-llava16/scripts/gates/llava_text_only/RUNBOOK.md)
in the training fork.