GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
adivaanihf
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
adivaanihf
Model Tree
sarvamai/sarvam-translate
Input Modalities
Output Modalities
Supported Functionality
adivaanihf/sarvam-bhili-to-marathi-loracheckpoint-17907 / epoch 3.0sarvamai/sarvam-translate37da6e80728634b661c3cc3ee2a3d9fa33ae9150ai4bharat/bhili-translate-bhb-mar435274ce4f40efc61dd1b355331f8edfeb007252sarvamai/sarvam-translateCAUSAL_LMr: 32q_proj, k_proj, v_proj, o_projgate_proj, up_proj, down_projfc1, fc2, out_projTraining used the source code included in this repository under source/.
BhiliMarathiadamw_torch_fusedThese are training-run metrics from trainer_state.json; they are not a final human evaluation.
checkpoint-17907checkpoint-11938 / epoch 2.0Important caveat: the training job entered a long/silent post-training evaluation/export stage after the above metrics. The adapter files here come from the completed latest checkpoint, and manual downstream evaluation is recommended before production use.
Main loadable adapter files at repo root:
adapter_model.safetensors — LoRA adapter weightsadapter_config.json — PEFT adapter configurationtokenizer.json, tokenizer_config.json, special_tokens_map.json, chat_template.jinja — tokenizer/chat assets copied from the checkpointtrainer_state.json — Hugging Face Trainer state and training logstraining_args.bin, scheduler.pt, rng_state.pth — training-state artifacts retained for provenance/resume diagnosticsAdditional files:
artifacts/training_summary.json — compact machine-readable training summaryartifacts/source_tree.txt — source file inventorysource/ — exact training/inference/evaluation source code and configs used for this runoptimizer.pt is intentionally not included to keep the model repo smaller; this repo is intended primarily for inference/evaluation from the adapter, not optimizer-level continuation.
bash
pip install -U transformers peft accelerate torch sentencepiece safetensors
Depending on your environment, you may also need any custom dependencies required by sarvamai/sarvam-translate.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelimport torchrepo_id = "adivaanihf/sarvam-bhili-to-marathi-lora"base_model = "sarvamai/sarvam-translate"# Use the same base revision used during training when exact reproducibility matters.base_revision = "37da6e80728634b661c3cc3ee2a3d9fa33ae9150"tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=False)base = AutoModelForCausalLM.from_pretrained(base_model,revision=base_revision,torch_dtype=torch.bfloat16,device_map="auto",trust_remote_code=False,)model = PeftModel.from_pretrained(base, repo_id)model.eval()
The training code formats translation prompts internally; for best reproducibility, inspect source/src and use the same prompt builder. A minimal direct example is:
python
text = "YOUR_BHILI_TEXT_HERE"prompt = f"Translate from Bhili to Marathi: {text}"inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.no_grad():output_ids = model.generate(**inputs,max_new_tokens=256,num_beams=4,do_sample=False,)print(tokenizer.decode(output_ids[0], skip_special_tokens=True))
For exact project-style usage, install the source package from source/ and use the included CLI/modules.
bash
git clone https://huggingface.co/adivaanihf/sarvam-bhili-to-marathi-loracd adivaanihf/sarvam-bhili-to-marathi-lora/sourcepip install -e .
The source tree contains:
configs/ — training configssrc/ — data loading, prompting, modeling, training, evaluation, inference, and pipeline codetests/ — unit tests used during developmentrun_training.sh / validate_data.sh — helper scriptsIf you use this adapter, cite:
sarvamai/sarvam-translateai4bharat/bhili-translate-bhb-maradivaanihf/sarvam-bhili-to-marathi-loraResearch, evaluation, and downstream low-resource MT experimentation for Bhili↔Marathi translation. Manual evaluation is recommended before using outputs in user-facing or high-stakes settings.