Professional card: Fireworks training, Ornstein thinking, RL and energy-based FT roadmap

README.md

@@ -11,38 +11,42 @@ pipeline_tag: image-text-to-text
 license: apache-2.0
 ---
 
+![Ornstein3.8-27B](ornstein3.8-27b.jpg)
+
 # Ornstein3.8-27B
 
-BF16 safetensors of **Ornstein3.8-27B** — a rank-32 LoRA trained on [Fireworks AI](https://fireworks.ai) and merged into [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B). Native vision-language model (`Qwen3_5ForConditionalGeneration`) with hybrid linear + full attention (Gated DeltaNet) and a 262K context window. Fine-tuning on Fireworks was straightforward.
+BF16 safetensors for **Ornstein3.8-27B**, a vision-language fine-tune of [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B). Architecture is `Qwen3_5ForConditionalGeneration`: interleaved linear and full attention (Gated DeltaNet), native image/video, 262K context.
 
-This checkpoint is meant to inject **Ornstein thinking** into Qwen3.8-27B. It is an early merge, not a finished quality release. Formal evals will be appended to this card as they land; further methods will drive the real quality gains.
+The LoRA was trained on [Fireworks AI](https://fireworks.ai) and merged into the Qwen3.8-27B language stack. Quantized GGUFs (Q8_0, Q6_K, Q4_K_M) and an mmproj are in [GestaltLabs/Ornstein3.8-27B-GGUF](https://huggingface.co/GestaltLabs/Ornstein3.8-27B-GGUF).
 
-GGUF quants (Q8_0 / Q6_K / Q4_K_M) plus an mmproj live in **[GestaltLabs/Ornstein3.8-27B-GGUF](https://huggingface.co/GestaltLabs/Ornstein3.8-27B-GGUF)**.
+## Status
 
-## Support This Work
+This checkpoint injects **Ornstein thinking** into Qwen3.8-27B. It is an early merge, not a finished quality release.
 
-I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.
+This card will be updated with formal evaluation as results land. Planned quality work uses **RL environments** and **energy-based fine-tuning**.
 
-**[Support on Ko-fi](https://ko-fi.com/djlougen)**
+## Support this work
 
----
+I'm a PhD student in visual neuroscience at the University of Toronto. Training and release compute is self-funded (rented H100s and a local DGX Spark). If these artifacts are useful, [Ko-fi](https://ko-fi.com/djlougen) helps keep the experiments running.
 
-## Model info
+## Model details
 
-- **Architecture:** `Qwen3_5ForConditionalGeneration` (linear + full attention interleaved, Gated DeltaNet; native image/video)
-- **Parameters:** ~27 B dense
-- **Context:** 262,144 tokens (native)
-- **Hidden size / layers:** 5120 / 64
-- **Attention:** 24 heads, 4 KV heads, head_dim 256
-- **MLP intermediate:** 17,408
-- **Vocab:** 248,320
-- **Precision:** bfloat16 safetensors (11 shards)
-- **Vision:** SigLIP-style tower, `out_hidden_size` 5120, patch 16
-- **Post-training:** PEFT LoRA rank 32 / α 32 trained on [Fireworks AI](https://fireworks.ai) (`ft-hk031mimxqpvk`), then merged into language-model linears only (vision + MTP left at base)
+| | |
+|---|---|
+| Architecture | `Qwen3_5ForConditionalGeneration` |
+| Parameters | ~27B dense |
+| Context | 262,144 tokens |
+| Hidden size / layers | 5120 / 64 |
+| Attention | 24 heads, 4 KV heads, head_dim 256 |
+| MLP intermediate | 17,408 |
+| Vocab | 248,320 |
+| Precision | bfloat16, 11 shards |
+| Vision | SigLIP-style tower, `out_hidden_size` 5120, patch 16 |
+| Post-training | PEFT LoRA rank 32, α 32, trained on [Fireworks AI](https://fireworks.ai); merged into language-model linears only (vision and MTP unchanged) |
 
 ## Usage
 
-Requires a recent Transformers with Qwen3.8 / `qwen3_5` support.
+Requires a Transformers build with Qwen3.8 / `qwen3_5` support.
 
 ```python
 from transformers import AutoProcessor, AutoModelForImageTextToText
@@ -64,15 +68,15 @@ messages = [
 ]
 inputs = processor.apply_chat_template(
     messages, add_generation_prompt=True, tokenize=True,
-    return_dict=True, return_tensors="pt"
+    return_dict=True, return_tensors="pt",
 ).to(model.device)
 out = model.generate(**inputs, max_new_tokens=256)
 print(processor.decode(out[0], skip_special_tokens=True))
 ```
 
-Text-only chat works the same way with `{"type": "text", ...}` and no image.
+Text-only chat uses the same template with `{"type": "text", ...}` and no image.
 
-vLLM / SGLang: load this repo as a Qwen3.8 27B VLM (`qwen3_5`). Pin a build that already supports that architecture.
+vLLM and SGLang: load this repo as a Qwen3.8 27B VLM (`qwen3_5`). Use a build that already supports that architecture.
 
 ## Files
 
@@ -82,8 +86,9 @@ vLLM / SGLang: load this repo as a Qwen3.8 27B VLM (`qwen3_5`). Pin a build that
 | `model.safetensors.index.json` | weight map, `total_size` 55562855904 |
 | `config.json` | `Qwen3_5ForConditionalGeneration` |
 | `tokenizer.json` / `tokenizer_config.json` / `vocab.json` / `merges.txt` | tokenizer |
-| `chat_template.jinja` | chat + vision + tool-call template |
+| `chat_template.jinja` | chat, vision, and tool-call template |
 | `preprocessor_config.json` / `video_preprocessor_config.json` | image/video processor |
+| `ornstein3.8-27b.jpg` | card banner |
 
 ## Related
 
@@ -93,4 +98,4 @@ vLLM / SGLang: load this repo as a Qwen3.8 27B VLM (`qwen3_5`). Pin a build that
 
 ## License
 
-Apache 2.0 — inherited from the Qwen 3.8 base release.
+Apache 2.0, inherited from the Qwen 3.8 base release.