VLR-CVC

CVC-Qwen3.5-35B-A3B-SFT

Deploy Dedicated

README

License: apache-2.0

Training

Two supervised finetuning stages, both at sequence length 16384 on 64 nodes of JUPITER (4x GH200 per node):

Instruct SFT — one pass over the LLaVA-OneVision-1.5-Instruct + Nemotron-VLM-V2 (197 sources), 2274 steps, 16.7B tokens
Decay anneal — a further 326 steps over the same sources reweighted towards document and OCR data, with the learning rate decayed to zero. ~2B tokens.

Benchmark Results

Evaluated with lmms-eval + vLLM on JUPITER (8 nodes, 32x GH200).

Table with columns: Benchmark, CVC-35B-A3B-Instruct, Qwen3.5-35B-A3B-Base, Qwen3.5-35B-A3B
Benchmark	CVC-35B-A3B-Instruct	Qwen3.5-35B-A3B-Base	Qwen3.5-35B-A3B
MMU_val (acc)	0.6833	0.6056	0.6267
MME (perception)	1769.7	1666.4	1560.8
MME (cognition)	762.9	740.4	751.8
AI2D (exact_match)	0.8598	0.0929	0.6758
OCRBench (accuracy)	0.8760	0.8920	0.9120
MMStar (average)	0.7373	0.7381	0.7289
ChartQA (relaxed_overall)	0.7720	0.2620	0.7572
MMLongBenchDoc (acc)	0.1790	0.2000	0.1980
MMLongBenchDoc (f1)	0.0360	0.0110	0.0060
DocVQA_val (anls)	0.9130	0.6515	0.9539
TextVQA_val (exact_match)	0.6573	0.2669	0.7665
InfoVQA_val (anls)	0.8570	0.3412	0.8622
SeedBench (seed_all)	0.7429	0.7479	0.7152
RealWorldQA (exact_match)	0.7438	0.6288	0.7739

Training data mixture

Both stages draw from the same 197-source LLaVA-OneVision-1.5-Instruct with Nemotron-VLM-Dataset-V2. The two stages differ only in how that blend is sampled. The exact list of subsets can be found on DATA_MIX.md.

Stable (instruct SFT) blend_epochized, every sample seen once (one deterministic full pass, ~16.7B tokens total). Composition follows each source's natural size:

Table with columns: Category, Share of stable-phase tokens, Example subsets
Category	Share of stable-phase tokens	Example subsets
OCR / scene-text (core)	1.34%	`textvqa`, `st_vqa`, `textocr_gpt4v`, `ocrvqa`, `llavar`, `textcaps`, `sroie_data`
Document VQA (core)	0.31%	`docvqa_train`, `infographic_vqa`

Decay (anneal) switched to a weighted blend (continuous sampling, ~2B-token budget) with four target categories deliberately upweighted:

Table with columns: Category, Share of decay-phase tokens
Category	Share of decay-phase tokens
OCR / scene-text (core, upweighted)	25%
Document VQA (core, upweighted)	15%
Spatial / real-world (core, upweighted)	12%
Chart / plot (core, upweighted)	8%
Text-only reasoning/math/code (replay)	13.79%
General captioning/instruct (replay)	9.97%
Nemotron-v2 text (replay)	8.10%
Document/OCR-adjacent, not core (replay)	6.72%

Notes

Uses the Qwen3.5 chat template. Unlike the upstream instruct template, this one renders <think> blocks on every assistant turn rather than only the last, matching how the model was trained.
eos_token is <|im_end|>, and generation_config.json stops on both <|im_end|> (248046) and <|endoftext|> (248044).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

VLR-CVC

Model Tree

Base

Qwen/Qwen3.5-35B-A3B-Base

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality