Spaces:
Sleeping
Sleeping
Commit History
-
Add full-sentence answer styling instruction
The full fine-tune's training data has many genuinely single-word or short-phrase ground-truth answers ("CMD", "E-8", a rupee figure), which the model absorbed as its default output shape - confirmed directly in production/eval samples ("D (P)", "1. HOD (HR) not below E-8"). Added an explicit instruction plus one example asking for complete-sentence answers instead of bare facts. Only one example is used, and deliberately not a monetary one: an earlier version of this instruction included both a role example and a monetary example ("instead of '₹10 lakh' write 'The limit is ₹10 lakh.'"), and validation showed the model anchoring on that literal figure - "₹10 lakh" started appearing as the answer to unrelated non-monetary questions (a grade question, a structural question) that it never produced before. Dropped the monetary example; the role-based one doesn't carry that anchoring risk. Validated on the same 20-question held-out set used throughout this work: no-context rate (5/20) and…e7e188f -
Fix production crash: empty-response error on ~8%% of real questions
Root cause: 'Answer:' was added to the llama.cpp stop-string list last push, intended to trim a redundant trailing restatement some completions produced. But this model also often *opens* its answer with the literal word 'Answer:' as a self-labeling habit - using it as a hard stop cut generation dead at position 0 in that case, producing a genuinely empty completion and surfacing as 'ValueError: Empty response from LLM' in production logs. Reproduced and measured directly: 5/60 (~8%%) of held-out questions crashed with this exact error before the fix. Fix, in two parts: 1. Removed 'Answer:' from the hard stop list. Added _clean_trailing_answer_restatement() instead, which distinguishes a genuine trailing restatement (real content before the marker - trim it) from the model's answer opening with 'Answer:' (nothing but whitespace before the marker - keep what follows instead of discarding everything). Applied to both the gguf and hf backends. 2. Generation is sampled (te…718a9e5 -
Deploy full fine-tuned TinyLlama and fix retrieval recall
Model: - Replace the LoRA-merged-and-quantized model with the full fine-tune from finetune_tinylama_full.ipynb, converted to GGUF and quantized to Q4_K_M (llama.cpp), committed via git-lfs as tinyllama_dop_full_q4_k_m.gguf (636MB). No separate model-repo upload needed; Dockerfile now COPYs it directly instead of curling it from a separate HF model repo at build time. - Added MODEL_BACKEND=hf as an additional (non-default) option to serve a raw transformers checkpoint instead of GGUF, for future local testing on GPU hardware; the deployed default stays MODEL_BACKEND=gguf, matching this Space's existing CPU-only setup. - Added PROMPT_TEMPLATE (default: inst), decoupled from MODEL_BACKEND. The full fine-tune was trained on Llama/Mistral-style "<s>[INST] ... [/INST]" pairs, not the old Zephyr-style <|system|>/<|user|>/<|assistant|> template this Space used for the LoRA model - prompt format now follows which model is served, not which library runs it. - Fixed a do…68e6652
-
improvements dec
0194a83
-
Update app/app.py
Verified signature98efe61 -
Update app/app.py
Verified signaturec560b3c
-
Update app/app.py
Verified signature52aa2ef -
Update app/app.py
Verified signature73fdab6 -
Update app/policy_vector_db.py
Verified signaturea179120 -
Update app/app.py
Verified signature85c3f24 -
Update app/app.py
Verified signaturef13ef99 -
Update app/app.py
Verified signature4dd3f6e -
Update app/policy_vector_db.py
Verified signature8193d78
-
Update app/policy_vector_db.py
Verified signature0c1cd5d -
Update app/policy_vector_db.py
Verified signature7118166 -
Update app/app.py
Verified signaturef3e86cb -
Update app/app.py
Verified signature7f343ef -
Update app/policy_vector_db.py
Verified signature784b064 -
Update app/app.py
Verified signature5340487 -
Update app/policy_vector_db.py
Verified signatureea6be63 -
Update app/app.py
Verified signaturef1d5824 -
Update app/app.py
Verified signatured5b1ff4 -
Update app/app.py
Verified signature9701d2e -
Update app/app.py
Verified signaturef220479 -
Update app/app.py
Verified signaturede7fbb1 -
Update app/app.py
Verified signatureacd510f -
Update app/policy_vector_db.py
Verified signaturef4341a1
-
Update app/app.py
Verified signature462ed44 -
Update app/app.py
Verified signature012e230 -
Update app/app.py
Verified signatured2159e2
-
Update app/app.py
Verified signatureb3975f9 -
Update app/app.py
Verified signature0a51966 -
Update app/app.py
Verified signature8ddf024 -
Update app/app.py
Verified signaturebd32824 -
Update app/policy_vector_db.py
Verified signaturece750f8 -
Update app/app.py
Verified signature7e6e5a8 -
Update app/app.py
Verified signatureb5cf3d3 -
Update app/policy_vector_db.py
Verified signaturef55e2f6 -
Update app/app.py
Verified signature1267728 -
Update app/app.py
Verified signature01687a9 -
Update app/app.py
Verified signaturebbdcb91 -
Update app/app.py
Verified signaturea00c3bb -
Update app/app.py
Verified signature4bf58b7
-
Update app/app.py
Verified signature4723486 -
Update app/app.py
Verified signature5043425 -
Update app/policy_vector_db.py
Verified signaturefbf0b84 -
Update app/policy_vector_db.py
Verified signature276e2ae -
Update app/app.py
Verified signature4d88d80 -
Update app/policy_vector_db.py
Verified signature994cab3
-
Update app/app.py
Verified signatureac06855