Downloads · 30 days
15
33% of all-time downloads
UlyssesXC/fab-qwen35-27b-lora-cot-127-epoch4-merged-20260924
fab-qwen35-27b-lora-cot-127-epoch4-merged-20260924 is a image-text-to-text model from UlyssesXC. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
Complete BF16 model weights with the epoch-4 LoRA adapter already merged. Load this repository directly; a separate adapter is not required. This is an experimental research checkpoint, not a validated financial advic…
Downloads · 30 days
15
33% of all-time downloads
All-time downloads
45
Public
Parameters
27.4B
54.7 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors54.7 GB · 100%
From the Hugging Face model README
Complete BF16 model weights with the epoch-4 LoRA adapter already merged. Load this repository directly; a separate adapter is not required. This is an experimental research checkpoint, not a validated financial advice system.
Base: Qwen/Qwen3.5-27B, revision fc05daec18b0a78c049392ed2e771dde82bdf654.
Training: 127 financial tool-use trajectories; prefixes end at complete assistant turns within a 49,152-token budget. Supervision includes normal assistant reasoning, text and tool calls; system/user/tool tokens and marked abnormal turns are masked. Four epochs, checkpoint step 128; LoRA rank 32, alpha 64, dropout 0, learning rate 5e-5, global batch 4, BF16 base. Language-model linear projections were adapted; vision parameters were frozen. PEFT merge_and_unload(safe_merge=True) produced the published weights. Original adapter SHA256: 14f04f79b3cad384223d91685b8549a60b9bc7c6eda04a7a0ab7a27682094676.
The files include the base tokenizer and chat template. Current experiment inference uses a separate LLaMA-Factory-compatible tool codec and removes past assistant reasoning while retaining current-turn thinking; loading the shipped standard chat template alone does not reproduce that history policy. Sampling: temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 1.5, repetition_penalty 1.0, seed 42. Context limit 262,144 including output; per-call output cap 32,768. No evaluation accuracy claim is made here.
Example download:
hf download UlyssesXC/fab-qwen35-27b-lora-cot-127-epoch4-merged-20260924 --local-dir ./fab-27b-epoch4
Example model serving (tested vLLM 0.27.1; provision sufficient GPU memory):
vllm serve ./fab-27b-epoch4 --dtype bfloat16 --language-model-only --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3
The example starts a standard model endpoint; the experiment-specific history and tool codec must be matched separately for comparable evaluation.